2022-03-17 20:14:02 RT @GoogleAI: Presenting PRIME, a data-driven approach for architecting hardware accelerators that trains a #DeepLearning model on existing… 2022-03-17 20:09:15 If you want to learn more about data-driven (offline) model-based optimization and COMs, you can also check out our earlier blog post on model-based optimization algorithms on the BAIR blog: https://t.co/E0d2ZNcstK https://t.co/t2iJmvFhqE 2022-03-17 20:05:58 Neural nets can design microchips from data to speed up neural nets! Check out @ayazdanb's &amp With offline model-based optimization based on COMs, PRIME can optimize accelerators from data. https://t.co/zXr2IIk9Db 2022-03-17 16:46:42 RT @hausman_k: We're releasing one of the biggest real-world multi-task RL robotic datasets! Our MT-Opt dataset has ~ 1M (!) RL trajectori… 2022-03-12 08:11:00 CAFIAC FIX 2022-01-17 08:11:00 CAFIAC FIX 2022-01-11 08:11:00 CAFIAC FIX 2021-12-27 08:20:00 CAFIAC FIX 2021-11-06 23:20:00 CAFIAC FIX 2021-12-18 23:00:05 @George_Mazz I sometimes think that good AI research really should make people at least *somewhat* uneasy. Obviously we should be thoughtful and responsible, but meaningful, important technological progress is always at least a little threatening. 2021-12-18 22:56:49 @UgandanDev Thanks. That part is also somewhat inspired by this rather interesting writeup by Gerald Loeb ("Optimal isn't good enough"): https://t.co/0HSc6RMNY5 2021-12-18 22:53:13 @ylecun @pmddomingos Thank you Yann! That means a lot to me. 2021-12-15 18:25:24 As usual, let me know if there is some paywall nonsense that medium starts to put up. I think it shouldn't happen, but if it does, I might migrate this over to a personal blog or something... 2021-12-15 18:24:48 I wrote an article about how we should situate our RL agents in realistic environments if we want them to acquire flexible behaviors: An Ecological Perspective on Reinforcement Learning Goes with the real-world RL talk &lt Article: https://t.co/pbPG9l3QhZ 2021-12-15 04:52:40 I'm posting my talk from the deep RL workshop here for anyone who didn't catch it: https://t.co/dFojc3ixts The Case for Real-World Reinforcement Learning Or: Why Robinson Crusoe didn't need to be good at chess in the hopes we can all keep it real when it comes to RL 2021-12-14 19:16:31 The full paper on Intrinsic Control via Information Capture (IC2) (Information is Power) is now on arxiv: https://t.co/ZWkhsDXOQk Find out how to implement "Maxwell's demon" as an entropy-minimizing self-supervised agent that hoards information &amp https://t.co/4gN62o3W17 https://t.co/U4SdAlUkQn 2021-12-14 05:06:52 Can we meta-train on one task at a time, and accelerate learning of each new task? Continual meta-policy search (CoMPS) enables online meta-RL, so that each new task learns faster by learning to learn https://t.co/yEeFuiPKhw w @GlenBerseth @WilliamZhang365 G. Zhang @chelseabfinn https://t.co/aQGcDBwKHT 2021-12-14 01:00:07 @ethancaballero @agarwl_ I don't think it fully fixes it. It mitigates it, but I think it's quite possible that with more research one could discover better methods to address this issue! 2021-12-13 21:31:52 My talk on why we should do RL in the real world is starting now at the deep RL workshop at @NeurIPSConf: https://t.co/18t15Op9ly 2021-12-13 18:57:35 This work was presented as a full-length oral at the NeurIPS Deep RL workshop: https://t.co/Is3skVO2mV Paper: https://t.co/esuL8Xfubf Work led by @aviral_kumar2, with @agarwl_, @tengyuma , @AaronCourville, @georgejtucker 2021-12-13 18:57:25 We can simply add DR3 to standard offline RL algorithms and boost their performance, basically without any other modification. We hope that further research on overparameterization in deep RL can shed more light about why deep RL is unstable and how we can fix it! https://t.co/VQhgh8dFg6 2021-12-13 18:57:13 In practice, we can simply add some *explicit* regularization on the features to counteract this nasty implicit regularizer. We call this DR3. It simply minimizes these feature dot products. Unlike normal regularizers, DR3 actually *increases* model capacity! https://t.co/B7ZBVgypsW 2021-12-13 18:57:00 But if we apply similar logic to deep RL as supervised learning, we can derive what the "implicit regularizer" for deep RL looks like. And it's not pretty. First term looks like the supervised one, but the second one blows up feature dot products, just like we see in practice! https://t.co/95tZrA4g3C 2021-12-13 18:56:46 When training with SGD, deep nets don't find just *any* solution, but a well regularized solution. SGD finds lower-norm solutions that then generalize well (see derived reg below). We might think that the same thing happens when training with RL, and hence deep RL will work well. https://t.co/GkEOO2fe3h 2021-12-13 18:56:31 Well, that's weird. Why do TD feature dot products grow and grow until the method gets unstable, while SARSA stays flat? To understand this, we must understand implicit regularization, which makes overparam models like deep nets avoid overfitting. 2021-12-13 18:56:17 Simple test: compare offline SARSA vs offline TD. TD uses the behavior policy, so same policy is used for the backup, but SARSA uses dataset actions, while TD samples *new* actions (from the same distr.!). Top plot is phi(s,a)*phi(s',a'): dot prod of current &amp 2021-12-13 18:55:51 Deep RL is hard: lots of hparam tuning, instability. Perhaps there is a reason for this? Turns out the same thing that makes supervised deep learning work well makes deep RL work poorly, leading to feature vectors that grow out of control: https://t.co/esuL8Xfubf Let me explain: https://t.co/f0EfM4wjGI 2021-12-13 02:22:40 WILDS, but now with unlabeled data! Representation learning from unlabeled data has the potential to be a powerful tool for improving model robustness, and WILDS v2.0 will hopefully enable further research on this. https://t.co/tpu0riVEWz 2021-12-09 19:42:17 @burgalon @NeurIPSConf @michaeljanner It's already out: https://t.co/98D21UVSMC 2021-12-09 04:34:33 Find out about how the principles behind CQL and MOPO can be combined into COMBO, a conservative model-based offline RL method! @TianheYu will present this tmrw 12/9 Thu at @NeurIPSConf 4:30 pm PT: https://t.co/xHf1vXA96q https://t.co/vU9JaarCuy 2021-12-09 04:18:03 Rewarding a policy w/ the accuracy of an underlying (latent space) model allows us to relate compression to reinforcement learning! @ben_eysenbach will present robust predictable control (RPC) @NeurIPSConf Thu 12/9 4:30 pm PT: https://t.co/QwuoMLRygK Spot: https://t.co/psTdS1HtTq https://t.co/JXbjNlXohP 2021-12-09 03:02:54 If you want to find out what kinds of mutual information maximization objectives are sufficient for RL, check out Kate Rakelly's @NeurIPSConf presentation tomorrow Thu 12/9: https://t.co/8vGfJ3Cnl9 Thu 12/9 4:30 pm PT! The paper is here: https://t.co/m2jCP1V6dL https://t.co/EkcezTPFUx 2021-12-08 21:09:28 Variational inference can be used to show that shaped goal-reaching objectives provide principled lower bounds on goal-reaching sparse rewards. Find out more at @timrudner and @Vitchyr's poster at @NeurIPSConf 8:30 am PT 12/9 Thu: https://t.co/JzF6Pth5oA Paper link https://t.co/hkCYH7D3fy 2021-12-08 21:06:46 RT @_prieuredesion: RECON was featured on Computer Vision News @RSIPvision: https://t.co/EFLMT9uzt2 I also presented RECON at @corl_conf i… 2021-12-08 05:25:42 Come find out how we can "compress" images so that user reactions (actions) in response to those images are preserved! Human-in-the-loop RL for compression with GANs. @sidgreddy will present PICO at @NeurIPSConf tomorrow at 4:30 pm PT: https://t.co/5YRx0ViEGf https://t.co/R9pNQx4jQM 2021-12-08 02:27:56 This makes RCE quite simple, and very effective at learning tasks without user-programmed reward functions, using only successful outcome examples. 2021-12-08 02:27:30 This problem statement is similar to VICE, which we've been using for many of our robotics experiments to learn rewards in the real world: https://t.co/0rLibIGN0H However, unlike VICE, there is only one model, which is both the classifier and the value function! 2021-12-08 02:26:43 The basic idea: instead of coding up a reward function by hand, provide a few example outcomes (states) that denote "success". RCE trains a classifier, which predicts whether an action will lead to "success" *in the future* https://t.co/f2cSfpLpYe 2021-12-08 02:25:32 Poster link: https://t.co/oCdbHOJgqr Oral link: https://t.co/UEGIXFjtA5 Blog post: https://t.co/H8Ri4uS0QI Paper: https://t.co/Djiv7RJzqY 2021-12-08 02:24:01 Classifiers can act as value functions, when a task is specified with a set of example outcomes. We call this method RCE. It's simple and works well! Find out more at @ben_eysenbach's poster Wed 12/8 4:30 pm at @NeurIPSConf, and full-length talk Fri 4:20 pm PT (oral sess 5)&gt 2021-12-08 02:19:33 RT @nick_rhinehart: Excited to present IC2, an agent that learns to discover and stabilize sources of uncertainty in partially-observed env… 2021-12-07 19:42:33 Find out how to "train on the test data" and get reliable likelihoods under distributional shift with BACS. Aurick Zhou will be presenting this work at @NeurIPSConf Wed 12/8 at 4:30 pm PT, poster session 4 poster C3! And you can check out the paper here: https://t.co/juFpiE6mb3 https://t.co/XucA8sKzCl 2021-12-07 19:11:33 Find out how Transformers can enable state-of-the-art offline RL results at @NeurIPSConf tomorrow! @michaeljanner will be presenting Trajectory Transformers 4:30 - 6:00 pm PT 12/8 (Wed): https://t.co/TZYIqRSwf4 Preview of the poster below. Paper is here: https://t.co/EiKXQGOPfx https://t.co/uxwfTmNx3Q https://t.co/0cJxKsHTWd 2021-12-07 04:46:54 IC2 will be presented at @NeurIPSConf by @nick_rhinehart tomorrow, Tue 12/7, at 4:30 pm PT in poster session 2, poster C0. You can check out the paper here: https://t.co/xZm2b4gDqO 2021-12-07 04:45:42 In the vizDoom video game environment, IC2 will look around to find enemies, and then shoot them, so that unpredictable enemies aren't there anymore (OK, this one is a bit violent... and maybe cause for some concern, but we'll find a way to apply it to more peaceful ends). https://t.co/vkN721AnXc 2021-12-07 04:44:27 For example, in a simple gridworld domain with moving objects that stop when the agent "tags" them, IC2 causes the agent to track down every object and tag it to stop its motion -- thus the agent always knows where everything is! https://t.co/57q02LUF4R 2021-12-07 04:43:27 Minimizing belief entropy forces the agent to do two things: (1) figure out where everything is (find &amp 2021-12-07 04:42:08 The idea behind IC2 (intrinsic control via information capture) is to instantiate this "belief entropy minimization" intuition into a practical unsupervised RL algorithm! There are a few variants of this principle, but they all train a latent belief model &amp 2021-12-07 04:40:49 This seems to violate the second law of thermodynamics. The explanation for why it does not is that information about the particles itself is exchangeable with potential energy (that's a gross oversimplifications, but this is just a tweet...). 2021-12-07 04:39:28 The "Maxwell's demon" thought exercise describes how information translates into energy. In one version, the "demon" opens a gate when a particle approaches from one side, but not the other, sorting them into one chamber (against the diffusion gradient). This lowers entropy. https://t.co/j8SKtceKAc 2021-12-07 04:37:56 Intrinsic motivation allows RL to find complex behaviors without hand-designed reward. What makes for a good objective? Information about the world can be translated into energy (or rather, work), so can an intrinsic objective accumulate information? That's the idea in IC2. A : https://t.co/H3xPmltZAI 2021-12-07 01:48:18 Come find out how the principles behind conservative Q-learning can facilitate selecting which data to share between tasks in multi-task RL! @TianheYu &amp 2021-12-06 22:47:19 If you want to find out why generalization in RL is hard, check out @its_dibya and @jrahme0's presentation at @NeurIPSConf tomorrow 12/7, 8:30 - 10:00 am PT poster session 1, poster C3! Meanwhile, check out @its_dibya's blog post on the epistemic POMDP: https://t.co/FBwQUrosQO https://t.co/Ieoc16DwLS 2021-12-06 22:42:00 In the real world, robots must train without someone resetting the world. Turns out that this can be an opportunity for a curriculum that can actually make learning easier! Check out Archit's presentation on VaPRL at @NeurIPSConf 8:30-10:00 am PT 12/7 poster sess 1, poster F1 https://t.co/P6yae6adwE 2021-12-06 22:38:46 Paper: https://t.co/u1FZBiUJZ1 Detailed blog post: https://t.co/3GReH3Z6vA 2021-12-06 22:38:23 The idea: construct unlabeled batches with different shifts at training time, and train the model to find the pattern (correlations) so as to do well on a labeled val set. Then at meta-test time, adapt to a new (unlabeled) batch. E.g., figure out the handwriting for a new user. https://t.co/E6cZ3ay6yt 2021-12-06 22:37:07 Can meta-learning be used to learn how to adapt to distributional shift *without* labels? On Tue 12/7 8:30-10:00 am PT, Marvin Zhang &amp poster session 1, poster A1 a short summary below: https://t.co/Aaaq2qtJPZ 2021-11-24 18:48:53 This leads to much better generalization for minigrid maze navigation and a complex image-based robotic manipulation environment. w/ @_prieuredesion, Peng Xu, Yao Lu, @xiao_ted, @alexttoshev, @brian_ichter https://t.co/Uh7A9trb5E 2021-11-24 18:46:31 This state can then be used with a high-level model based planner, or a high-level model-free RL method. Insofar as the low-level skills can generalize, the high-level policy has a much easier problem to solve, because VFS abstracts away perception. https://t.co/6fuVVqJZuR 2021-11-24 18:45:45 The idea is simple: if we have a bunch of low-level skills, each skill contributes its value function as one entry in a high-level state. So if we have skills for opening a door and opening a drawer, higher level sees the state as "can open door, can't open drawer." 2021-11-24 18:44:28 Value function spaces (VFS) uses low-level primitives to form a state representation in terms of their "affordances" - the value functions of the primitives serve as the state. This turns out to really improve generalization in hierarchical RL! https://t.co/yJqJwwCT6r Short &gt 2021-11-23 18:34:24 Which MI rep learning objectives are sufficient for RL? Kate Rakelly analyzes this in her new blog post https://t.co/L8PteaDk7y If you use an obj that is insufficient, you may get a representation that can't solve your task. Based on our NeurIPS paper https://t.co/Wg1vUe8Ouh https://t.co/oWI9xLSjbO 2021-11-22 20:52:27 An awesome opportunity for students who want to work on RL, robotics, and other cool topics at MILA :) https://t.co/gg8tTgX5vz 2021-11-22 18:52:07 The full bridge dataset that we are releasing uses a cheap widowX arm to collect 7,200 demonstrations for 71 tasks across 10 different domains (themed around household tasks). Find dataset, code, and paper link on the project website: https://t.co/fGG35ikRus 2021-11-22 18:50:46 However, if it has enough *different* tasks and domains, this turns out to actually work! In fact, including bridge data when training new tasks in new domains (in this case with imitation learning) leads to about a 2x improvement in performance on average! https://t.co/n6cmw22Uwb 2021-11-22 18:50:03 So if I collect a little bit of data for my task in my new domain, can I use a reusable dataset to boost generalization of this task? This is not a trivial question, since the "bridge data" does not contain either the new domain or the new task. https://t.co/876qCH8aKf 2021-11-22 18:49:22 The reason reusing data in robotics is hard is that everyone has a different lab environment, different tasks, etc. So to be reusable, the dataset needs to be both multi-domain *and* multi-task, so that it enables generalization across domains. 2021-11-22 18:48:23 Reusable datasets, such as ImageNet, are a driving force in ML. But how can we reuse data in robotics? In his new blog post, Frederik Ebert talks about "bridge data": multi-domain and multi-task datasets that boost generalization of new tasks: https://t.co/JbIbC9X1I1 A thread: 2021-11-19 20:01:51 Also, if you want to read the paper that we "borrowed" the Q-function from for Ant Maze, it's here: https://t.co/pksjjvvCkH @ikostrikov makes some really nice Q-functions 2021-11-19 19:58:46 This is joint work with Michael Janner &amp https://t.co/98D21Vdubc https://t.co/EiKXQGxenZ Code: https://t.co/gG30OKwjgJ 2021-11-19 19:58:34 This is significant because only dynamic programming methods perform well on Ant Maze (e.g., Decision Transformer is on par with simple behavioral cloning) -- to our knowledge Trajectory Transformer + IQL is the first model-based approach that improves over pure DP on these tasks 2021-11-19 19:58:13 But if we *combine* Trajectory Transformer with a good Q-function (e.g., from IQL), we can solve the much more challenging Ant Maze tasks with state-of-the-art results, much better than all prior methods. Ant Maze is much harder, because it requires temporal compositionality. https://t.co/OzNjmbM37X 2021-11-19 19:57:24 For control, we can simply run beam search, using reward instead of likelihood as the score. Of course, we could use other planners too. On the (comparatively easy) D4RL locomotion tasks, Trajectory Transformer is on par with the best prior method (CQL). https://t.co/JkWJx6oDWN 2021-11-19 19:57:07 It also makes *very* long-horizon rollouts successfully, far longer than standard autoregressive models p(s'|s,a). So something about a big "dumb" model works very well for modeling complex dynamics, suggesting it might work very well for model-based RL. https://t.co/b1zj5O4HYZ 2021-11-19 19:56:10 Although the Transformer is "monolithic", it *discovers* things like near-Markovian attention patterns (left) and a kind of "action smoothing" (right), where sequential actions are correlated to each other. So the Transformer learns about the structure in RL, to a degree. https://t.co/WPIpajbmFs 2021-11-19 19:55:53 Trajectory transformer is a "one big dumb model" approach to model-based RL: every single dimension of every state and action is a (discrete) token in a huge sequence. The model doesn't distinguish between states vs actions vs rewards, they're all just tokens. https://t.co/d1Vpz1YWgl 2021-11-19 19:55:39 We've updated Trajectory Transformer (Transformers + model-based RL) for NeurIPS camera-ready, now with more complete results that include Ant Maze, including value functions, and also a blog post summarizing the method! https://t.co/EiKXQGxenZ https://t.co/xoOMQznKLW A thread: 2021-11-17 04:56:44 @CsabaSzepesvari @daibond_alpha @nanjiang_cs Good motivation for another reunion workshop next year? 2021-11-12 02:10:52 AW-Opt combines AWAC and QT-Opt into a scalable algorithm for real-world offline RL with online finetuning, particularly for combining imitation learning and RL. Evaluated on multiple complex tasks from our work on large-scale robotic learning at Google! https://t.co/CmPDowzQNH 2021-11-10 20:01:31 I'll be presenting (and discussing) this position paper at @corl_conf tomorrow (Thu 11/11) at 6:30 am PT (sorry...), which is 14:30 GMT! See you there. Meanwhile, here are a couple slides from my talk to give you some idea of what I'm going for. https://t.co/yRDYdeMAVY https://t.co/wdrLOTordg 2021-11-10 19:56:00 Check out Natasha's tutorial tomorrow at @corl_conf, which will cover cool work on social learning, emergent complexity, and other topics! https://t.co/twAVYEc9d1 2021-11-10 18:50:05 On Thu 11/11, @katie_kang_ will present this work at @corl_conf! Poster session at 5 pm GMT (9 am PT). Come find out how different robots (and even humans!) can share data to learn a single unified navigational system with different per-robot dynamics. https://t.co/HLTTjJD11S https://t.co/iSzgtVTQIb 2021-11-10 18:47:04 Want to know how to get a robot to learn to clean up your room? Thu 11/11, we'll be presenting ReLMM at @corl_conf, poster 11:30 am GMT. Find out about autonomous real-world robotic RL that combines mobility + manipulation! https://t.co/RiPiTFkLL0 https://t.co/qYQJJbWVni 2021-11-09 18:12:26 How do we tune offline RL methods? In our full-length @corl_conf talk, @aviral_kumar2 &amp 2021-11-09 17:02:35 @pcastr @its_dibya The principle actually applies to any generalization across anything, but it’s easiest to describe across contexts in a contextual MDP, because this makes it simplest to define a train/test split. But the same thing happens if we have unseen (but in dist) states at test time 2021-11-08 18:44:17 MT-Opt will be presented at @corl_conf Tue 11/9 in the 11:30 am GMT poster session! Meanwhile, you can find out more about how large-scale robotic deep RL can learn diverse tasks here: https://t.co/yPQ7wYFjM5 Paper: https://t.co/gweqaiT7oC Video: https://t.co/R3Wg6TYnaP https://t.co/jH6qPBdxbT 2021-11-08 18:42:25 We'll present RECON in the 10 am GMT oral session on Tue 11/9 at @corl_conf (poster 11:30 am). If you want to get an early preview, you can find a recorded talk here https://t.co/tvCAXZvhcs Come find out how robots can learn to search for new goals! Paper https://t.co/WsEYRIb6P9 https://t.co/PP9csCOijU 2021-11-07 22:30:51 @minakhan01 That's very kind of you to say, thank you This also reminded me that I really need to prep for my class discussion tomorrow. 2021-11-07 20:53:06 Also, sorry about the slightly awkward format with the giant face taking up half the screen, that was the format I recorded it in that was requested by the organizers But I guess it's nice if you prefer to see the speaker in a talk... 2021-11-07 20:52:13 I'm reposting my BayLearn 2021 keynote talk on offline RL here: https://t.co/YgCyeFknzz Thank you to the BayLearn organizing committee for the invitation, it was a lot of fun. This talk covers CQL, IQL, and other algorithms &amp 2021-11-07 17:40:25 @sanjeevanahilan Conceptually quite related! We discuss this in the related work section of our NeurIPS paper (just wanted to keep the blog post brief and accessible). 2021-11-06 23:20:00 CAFIAC FIX 2021-11-06 19:50:00 CAFIAC FIX 2021-11-06 18:59:00 CAFIAC FIX 2021-11-01 19:20:00 CAFIAC FIX 2021-11-01 17:30:00 CAFIAC FIX 2021-10-13 19:14:28 RT @tonyzzhao: "Insert anything into anything!" New paper applying offline RL to industrial insertion. Test it with 12 new tasks, 100/100… 2021-10-13 18:41:06 RT @ikostrikov: Excited to present our work with @ashvinair and @svlevine, Offline RL with Implicit Q-Learning (IQL), a simple method that… 2021-10-13 18:30:10 Now we'll see if we can use IQL for our robotics applications. So far seems like a great choice, as it's very fast, performs great, and finetunes really well. 2021-10-13 18:29:33 This leads to state-of-the-art results on D4RL offline RL benchmarks, with particularly large gains on the hardest task (e.g., large ant mazes). Also excellent online finetuning results. And really fast runtime. https://t.co/6vwifK12Qv 2021-10-13 18:29:18 Just by changing the loss function for V(s) like this, we can get V(s) to converge to the maximum *in-support* value function, resulting in honest-to-goodness dynamic programming! This allows learning the optimal policy in the data support. https://t.co/SPgwP4Z6T4 2021-10-13 18:29:06 Now the kicker: change the loss for V(s) 2021-10-13 18:28:47 In general, we could train another network to be the value function V(s), and then use a loss like the one below. If V(s) is trained by regressing onto Q(s,a) using dataset actions (remember, we *only* use dataset actions), it's *still* SARSA, no improvement. https://t.co/CXzS8Ev9cH 2021-10-13 18:28:30 Here is the idea: if we want to prevent *all* OOD action issues in offline RL, we could use *only* actions in the dataset. That leads to a SARSA update, which is very stable. But it learns the *behavior policy* value function, not the optimal value function: https://t.co/6omLFHuHWb 2021-10-13 18:28:17 Implicit Q-learning, or "I can't believe it's not SARSA": state-of-the-art offline RL results, fast and easy to implement w/ @ikostrikov, @ashvinair -&gt 2021-10-13 03:58:33 If you want to learn more about the algorithms that went into this, check out: PEARL (meta-learning with probabilistic encoder): https://t.co/Ebiwsdg4nH AWAC (actor-critic offline RL): https://t.co/913iRX0ppL SMAC (offline meta-RL algorithm): https://t.co/McmMFHkzJ8 2021-10-13 03:57:27 Even works on some super-hard challenge tasks, like RAM insertion. Find out more on the project website: https://t.co/XIjDBtfhwx arxiv: https://t.co/LOqgMH5Z48 video: https://t.co/zUbHMWgHu9 https://t.co/GGs73yVmTS 2021-10-13 03:57:04 For the connectors that don't work well, run additional online finetuning, which gets their performance up to 100% in about 5-10 minutes! https://t.co/0iezSp2Ugu 2021-10-13 03:56:09 The method: use 11 meta-training tasks to meta-train a PEARL-style encoder using AWAC. Use this to then adapt to new connectors: some work well, some don't. https://t.co/L1vJpNHeYY 2021-10-13 03:55:47 Offline RL + meta-learning enables industrial robots to learn new insertion tasks with near-perfect success rates with AWAC + PEARL + finetuning! w/ @tonyzzhao, @jianlanluo, @DeepMind, Intrinsic https://t.co/XIjDBtfhwx Short summary below: https://t.co/TjYuftKpwj 2021-10-12 17:45:46 For more, check out: Paper: https://t.co/66ztHfpms3 Website: https://t.co/TkJw1eZ4dJ Video: https://t.co/bly0VuzQAa w/ Laura Smith, Chase Kew, @xbpeng4, Sehoon Ha, Jie Tan. An awesome collaboration between @berkeley_ai and @GoogleAI! 2021-10-12 17:43:46 Why is this important? We can never pretrain robots so that they never fail. But with RL, the robot can just keep on learning, and adjust to recover from failure even in unexpected conditions. This is going to be crucial for real-world robustness. https://t.co/ufdtkAT5CL 2021-10-12 17:42:25 The main idea in our work is to use RL to finetune walking policies to the particular environment where they are deployed. We learn policies for going forward and backward with motion imitation and a highly agile reset controller, which enable automated real-world finetuning. https://t.co/wbWubTtEql 2021-10-12 17:42:05 Can walking robots be made robust enough to finetune with RL directly in the real world? In "Legged Robots that Keep on Learning", we explore this question, using learned highly agile reset controllers and sample-efficient RL: https://t.co/TkJw1eZ4dJ https://t.co/66ztHfpms3 -&gt 2021-10-12 04:00:53 RT @chelseabfinn: Robot learning is bottlenecked on good, reusable datasets We introduce: * a new dataset with demos of 71 tasks over 10 e… 2021-10-11 20:51:34 @maxhkw We tried ImageNet in prior projects. It can help a little, but not a lot. My sense is that visual features are only a modest part of the issue. We need to analyze what is transferred from bridge data more, but intuitively it seems like reaching toward objects is a part of it... 2021-10-11 17:34:10 Here is another bonus animation of more cool robot behaviors https://t.co/SSOPhAjw4a 2021-10-11 17:33:05 The dataset we collected consists of 7,200 demonstrations, over 71 tasks, and 10 different domains. This project was led by @febert8888 with many great collaborators! Paper: https://t.co/25guMobOVS Website w/ videos, data, code: https://t.co/fGG35ikRus 2021-10-11 17:32:55 Experimentally, this works great: blue is with bridge data, red is without. All tasks are unseen in the bridge data, and the environment is *also* unseen. Just including bridge data (containing other tasks and other environments) boosts generalization by 2x! https://t.co/athpGqQLan 2021-10-11 17:32:37 The point of "Bridge Data" is to serve as a large task-agnostic dataset that can be included (as pretraining or joint training) to boost generalization for a new task in a new domain. So everyone uses the same bridge data, and small in-domain datasets in their lab, for their task https://t.co/qCHYGoHqOl 2021-10-11 17:32:18 In robotic learning, we start from scratch for every experiment, with custom per-task data. Generalization is poor, because the dataset is narrow. In supervised learning domains, there are large reusable datasets (e.g., ImageNet, MS-COCO). What would be "ImageNet" in robotics? https://t.co/qyC9vfac5C 2021-10-11 17:31:54 When training your robot, don't start from scratch! In "Bridge Data", we study how big datasets robotic datasets (71 tasks, 7.2k demos!) bridge the generalization gap. For new tasks, use bridge data along with task data to boost generalization: https://t.co/fGG35ikRus A thread: https://t.co/ed8m1hS95s 2021-10-08 04:09:50 This leads to better accuracy *and* better calibration on unlabeled test points. Inspired by some classics on entropy minimization: Y. Grandvalet, Y. Bengio. Semi-supervised Learning by Entropy Minimization M. Seeger. Input-dependent Regularization of Conditional Density Models 2021-10-08 04:09:06 To avoid needing to store all training data, we can learn posterior q(theta) using any BNN approach, and then incorporate this as a regularizer when minimizing entropy at test time on unlabeled data. https://t.co/cHlCQLDQRu 2021-10-08 04:08:52 This naturally leads to an entropy minimization procedure at test time: get some unlabeled points, and then update the parameter posterior to get lower entropy on test points, but don't stray too far from parameter distribution on training set! https://t.co/TdgRlHmEZQ 2021-10-08 04:07:50 We need a better graphical model. What if we assume new datapoints are not arbitrary: if we are asked to classify a new OOD, it likely belongs to *one of* the classes, we just don't know which one! Now there is a relationship between theta and phi for each distribution! https://t.co/2BPaQCR0yh 2021-10-08 04:07:35 First: what information do we gain from observing unlabeled datapoints from a new distribution? We can draw a graphical model for this: x is input, y is label, theta is classifier params, phi parameterizes x distro. Unfortunately, if y is unobserved, x tells nothing about theta! https://t.co/E2w5hxaYyV 2021-10-08 04:06:49 Is there a principled way to adapt a model to distributional shift without labels? In our new paper, "Training on Test Data", we propose a Bayesian adaptation strategy based on BNNs and entropy minimization. w/ Aurick Zhou: https://t.co/juFpiE6mb3 A thread: 2021-10-07 20:52:02 RT @ben_eysenbach: Unsupervised skill learning methods based on mutual information (e.g., DIAYN) learn a wide range of skills, which are of… 2021-10-07 20:35:03 This could be viewed as a kind of good initialization for downstream policy finetuning, formally it optimizes this objective. https://t.co/sB93xNC0Oz 2021-10-07 20:34:46 Turns out they don't. It means no matter how many skills you try to learn, you will not capture all possible optimal policies! But there is some good news -- what you do learn is a group of skills whose average minimizes distance to the "hardest" (furthest) policy. https://t.co/3vYP2cPAuH 2021-10-07 20:33:39 RL objective is linear in state marginal space, which means optimal policies must be vertices of the orange polytope. Do unsupervised RL methods get all vertices (all optimal policies)? 2021-10-07 20:33:28 We can use a geometric interpretation of RL policies: the triangle represents a polytope over state probabilities (3 states here). Orange region is states you can reach. https://t.co/lf9QP2uZ0s 2021-10-07 20:33:14 For a while, @ben_eysenbach, @rsalakhu, and I have been trying to understand unsupervised RL, such as DIAYN https://t.co/ZH51PsSkqj Which skills do they learn? Can we develop a theory of unsupervised RL? Our paper https://t.co/zst5S8K3CC Thread on what we found (and that gif): https://t.co/dz2u1tnUC1 2021-10-07 18:29:46 Generative models can be used for compression. But if we combine them with an "RL-like" GAN, we can make it so that, when a user sees the compressed image, they do the same thing as if they saw the original. This is the idea behind pragmatic compression: https://t.co/N2wFFyV0uw 2021-10-07 04:23:13 The model is still tilted toward optimistic outcomes, but this actually aids in exploration, leading to faster learning. See more analysis in the paper: https://t.co/GII1kNcNiN Code for a toy example here: https://t.co/bflWYjhrG4 w/ @ben_eysenbach, @AlexKhazatsky, @rsalakhu 2021-10-07 04:20:51 The model optimizes the *same* objective: maximize reward (value function) and also fool the discriminator, thus producing trajectories that are realistic and have high reward. The discriminator punishes the model for cheating (going to high reward regions unrealistically). https://t.co/vr0FEfP6W4 2021-10-07 04:19:48 The policy objective is to maximize reward and also minimize difference between learned and true dynamics *on the states the policy visits*. This can be quantified w/ a discriminator -- like the generator in a GAN, the policy seeks out "realistic" trajectories (with high reward). https://t.co/AHWp01IiiK 2021-10-07 04:18:03 MnM instead optimizes a *lower bound* on the policy reward with respect to both the model and the policy, using a GAN-like model formulation. While this might seem a bit counterintuitive, it leads to an algorithm that provably improves a bound on the reward. https://t.co/AfpVq2ZfVe 2021-10-07 04:16:36 Regular model-based RL methods train the model with MLE, but better models don't necessarily translate to better policies, hence understanding convergence is difficult. Can we devise a MBRL method where the model optimizes an objective that makes it better *for the policy*? https://t.co/0EfpOa0bmD 2021-10-07 04:14:53 Can we devise model-based RL methods that use the *same* objective for the model and the policy? The idea in Mismatched no More (MnM) is to devise a single objective that can be optimized by both model and policy that provably lower bounds reward. A thread: https://t.co/zxFy7nfpii 2021-09-27 17:08:07 How do we get robots to learn what they *can* do from prior data, so that they can learn what they *should* do via RL? Visual affordance learning (VAL) does this via goal-conditioned RL and generative models. New blog post by @ashvinair &amp https://t.co/S1H0UnKy1X https://t.co/OkgBoq73FD 2021-09-23 17:09:48 This paper will be presented at CoRL 2021, with @aviral_kumar2, Anikait Singh, Stephen Tian, @chelseabfinn https://t.co/Yg9YDMrIA0 https://t.co/WqeCOQeGQ5 2021-09-23 17:08:03 A few things that I think are interesting: (1) we can do capacity/arch/hyperparam tuning *without* full OPE (which is very hard) 2021-09-23 17:06:29 We evaluate these guidelines on a simulated robotic task, and two different real-world robots, and find that it works well across the board, using the same alpha=1.0 CQL parameter and fully offline selection of model capacity, regularization, etc. https://t.co/zeCp1Gdhzt 2021-09-23 17:05:19 Of course, the true return is unknown during offline training, but we can still use our understanding of the trends of estimated Q-values to provide guidelines for how to adjust model capacity. These guidelines are not guaranteed to work, but seem to work well in practice. https://t.co/TTaI5x2iwP 2021-09-23 17:02:17 In supervised learning, we have training and validation sets, and this works great for tuning. There is no equivalent in RL, making tuning hard. However, with CQL, when there is too much or too little model capacity, we do get very characteristic estimated vs true Q-value curves https://t.co/s5YE6O7Ger 2021-09-23 17:00:42 Offline RL lets us run RL without active online interaction, but tuning hyperparameters, model capacity etc. still requires rollouts, or validation tasks. In our new paper, we propose guidelines for *fully offline* tuning for algorithms like CQL https://t.co/Yg9YDMrIA0 A thread: https://t.co/ibqI3X7FDG 2021-09-17 20:57:18 RT @hausman_k: When working on MT-Opt, we saw that sharing data in multi-task RL can make or break the algorithm. Conservative data shari… 2021-09-17 16:54:21 CDS was developed by @TianheYu &amp 2021-09-17 16:52:14 In practice, CDS is quite easy to implement, using the Q-function from CQL to estimate what to share, and leads to significant improvements across a range of multi-task offline RL problems. https://t.co/rRREPShlrE 2021-09-17 16:51:05 At the same time, sharing as much as possible is good, because it gives you more samples. This tradeoff between sharing more to get samples and sharing less to get distributional shift is captured by our main theoretical result, which formalizes this tradeoff. https://t.co/nMY5oqJmwr 2021-09-17 16:50:06 Now it's clear we should choose what to share so that pi_Beff is close to what we think a good policy pi should be, to minimize the second term (the divergence). This intuitively means: share from tasks that are similar to how you *think* the new current task should be done! 2021-09-17 16:49:22 The induced pi_Beff is the behavior policy you constrain too, and you can't go too far from pi_Beff (else the Q-function diverges). But choosing what to share *changes* pi_Beff. Imagine that offline RL optimizes an objective of this form (turns out CQL does this more or less): https://t.co/cFOG90HNqM 2021-09-17 16:47:35 More technically, imagine "pi_Beff" is the "effective" policy induced by data sharing for task B, which is a mixture of the behavioral policy for task A and task B. For offline RL methods that are conservative (necessary to avoid divergence!), pi_Beff is very important. 2021-09-17 16:45:08 We first hypothesize that the reason sharing data across tasks (i.e., using data for task A to help task B) makes offline RL hard is that it increases distributional shift: if task A is irrelevant, we're adding data from a very different distribution. 2021-09-17 16:44:13 Multi-task RL is hard. Multi-task offline RL is also hard. Weirdly, sharing data for all tasks (and relabeling) can actually make it harder. In conservative data sharing (CDS), we use conservative Q-learning principles to address this. arXiv: https://t.co/4hhl639bgh A thread: https://t.co/wo3ZLrj52A 2021-09-17 16:40:37 RT @chelseabfinn: In offline RL, adding data from other tasks can boost generalization but can surprisingly *hurt* performance We analyze… 2021-09-17 15:30:40 RT @gradientpub: Check out our interview with @berkeley_ai professor @svlevine! We talk about his start in research, DiscoRL, sequence mod… 2021-09-16 17:29:26 A fun discussion with Andrey on robotic learning, RL, and where the arm farm came from. https://t.co/Q2oZsMW2az 2021-09-14 05:07:10 RT @_prieuredesion: RECON accepted as an Oral Talk at @corl_conf 2021!! What are the odds we actually get a live audience in London? 2021-09-12 17:23:24 @Shixo I suspect that is right, but it’s not intended to improve exploration but rather final robustness and generalization (via compression). I suspect that, like SMiRL, it can be combined with novelty seeking exploration in the obvious way (it might seem contradictory but it’s not) 2021-09-10 16:47:00 @TongzhouWang It's not a generative model of states 2021-09-10 16:45:56 @pcastr The policies trade off information and reward, so fully open-loop policies are definitely not going to be optimal in general, but the goal is just to get them to minimize unnecessary input (that's of course an old idea, the difference is in the approach) 2021-09-10 03:14:51 @Julius_Frost The policy still gets a standard task reward -- all this stuff is just added on top of regular RL. 2021-09-10 02:13:50 This work is by Ben Eysenbach, w/ @rsalakhu Paper: https://t.co/xIgZPiKiKr Website with short video: https://t.co/NmR8Aa8EaU Code: https://t.co/MnNpeSXiTB 2021-09-10 02:13:39 Empirically, this leads to much better policies if we want low bitrates, more robust behavior to sensor dropout and perturbations, and more compressed representations that help generalization! https://t.co/A9LKoXz9dz 2021-09-10 02:13:27 Theoretically, one nice consequence of this is that, if we use a big enough multiplier on the accuracy term, we train a policy that tries to be as good as possible in open loop (see theorem). https://t.co/VafSpzDZcY 2021-09-10 02:13:10 This means the policy, encoder, and model "collude" to find a z that is minimal (easiest to predict) but still enough to take good actions. This is exactly the same as compressing the state to the smallest number of bits needed to get high reward! 2021-09-10 02:12:57 This means we train an encoder phi(z | s), a policy pi(a | z), and a model m(z[t+1]|z[t],a[t]). The only thing that makes z non-trivial is that it must decode to actions for the policy, so it only contains the info necessary to take actions. https://t.co/vFN00M9keT 2021-09-10 02:12:42 The idea is simple: learn a model m(z[t+1]|z[t],a[t]) that predicts a future latent state, and reward the policy for taking transitions for which the predicted latent state matches what we get by encoding the *actual* next latent state. https://t.co/vNJnxHkZuV 2021-09-10 02:12:26 What happens if, when we learn a model in RL, we reward the policy for doing things where the model is correct? Turns out that this is (surprisingly) the same as minimizing the number of bits of state that the policy needs access to. Paper: https://t.co/KIDQOM3iaW Thread below: https://t.co/9HBbJ9GrnM 2021-08-17 16:05:01 RT @SeungmoonS: "Deep reinforcement learning for modeling human locomotion control in neuromechanical simulation" is now published at JNER:… 2021-08-12 07:57:06 RT @xbpeng4: We will be presenting our work on adversarial motion priors tomorrow 4pm PST at #SIGGRAPH2021. Drop by the session if you want… 2021-08-09 19:29:21 @pmddomingos But can one get a Nobel Prize for 280 characters? 2021-08-08 22:51:49 @KuanFang Congratulations Kuan! Looking forward to working with you soon 2021-08-06 17:17:43 RT @ColinearDevin: So excited to release our work on real world autonomous learning! https://t.co/CpgZrp5Zhk 2021-08-06 03:26:05 RT @GlenBerseth: New work on fully autonomous real-world robotic learning with mobile manipulators. One of the exciting parts of this work… 2021-08-05 03:30:18 RT @abhishekunique7: New work on learning how to grasp and navigate with mobile robots using RL. What I find very exciting is the ability o… 2021-08-05 03:24:59 Arxiv: https://t.co/RiPiTFkLL0 Website with video: https://t.co/DNSnELotdX w/ Charles Sun, @jendk3r, @ColinearDevin, @abhishekunique7, @GlenBerseth 2021-08-04 20:30:22 Arxiv: https://t.co/RiPiTFkLL0 Website with video: https://t.co/DNSnELotdX w/ Charles Sun, Jedrzej Orbik, Coline Devin, Brian Yang, Abhishek Gupta, Glen Berseth 2021-08-04 20:30:12 The more ReLMM trains, the better it gets. It can train for 60+ hours autonomously (with just battery changes), and performance just keeps getting better and better, as shown below. https://t.co/RacLGhjPFB 2021-08-04 20:29:58 The cool thing is that we can run it autonomously in many different rooms, with the only human effort being to change batteries. https://t.co/0j99Jqtxdo 2021-08-04 20:29:46 ReLMM makes it possible for robots to learn continually and without resets. In our evaluation, the task is to clean up a room, picking up every object and depositing it in a basket. ReLMM uses a grasping and navigation policy, with uncertainty from grasping providing exploration. https://t.co/ZuvB4Ly9Ro 2021-08-04 20:29:30 Robots with RL can learn autonomously in the real world, and this allows them to keep getting better and better. We explore this in our recent paper, introducing ReLMM, a system for fully autonomous real-world robotic with mobile manipulators: https://t.co/DNSnELotdX A thread: https://t.co/PCGdXStDUp 2021-08-04 03:23:17 RT @HaoSuLabUCSD: Excited to announce the SAPIEN Manipulation Skill challenge, a learning-from-demonstrations challenge from 3D visual inpu… 2021-07-30 04:23:54 @Brugzstopher One way to think of it is that it's just like regular meta-RL during this step, only the rewards for the tasks are not real rewards, but produced by the model. This still works, because those tasks just need to be varied and representative of downstream tasks for meta-training. 2021-07-29 20:25:14 When RL learns in the real world, it can't push a "reset button" to try again, but must do lifelong reset-free learning. Maybe harder, but also easier, because it gives an opportunity to automatically build a curriculum! We leverage this in our new paper: https://t.co/TdsAmp8XPw https://t.co/P6yae6adwE 2021-07-29 18:57:11 @pfau I prefer "intellectrician". Feels more apt sometimes... 2021-07-29 04:17:30 RT @hausman_k: Resets are one of the most limiting, often under-emphasized requirements of current robotic RL methods. They are hard to aut… 2021-07-28 17:19:47 And of course I should add that theoretical RL research *has* been examining environment properties through assumptions (e.g., linear MDP), but the empirical RL research community has given less attention to this, preferring standardized benchmarks. 2021-07-28 17:17:30 This is almost obvious: we know in supervised learning data is just as critical as algorithm we know that what animals learn is extremely dependent on their environment Likely impossible to get "general-purpose" intelligent agents unless they are situated in a proper environment 2021-07-28 17:15:44 I'm excited to co-organize this WS on a unique topic. The "ecological" perspective on RL is one that examines the role of environment, task, data, etc. rather than just the algorithm, and I think this is really important. We had an initial paper on this: https://t.co/35ibxthqyY https://t.co/SII4rcDyFH 2021-07-27 17:16:27 We evaluated it on some fun robotic manipulation tasks, where we test generalization to entirely new tasks with similar objects. These are based on tasks from the VAL paper: https://t.co/pf7qqtM8i1 w/ @Vitchyr @ashvinair Laura Smith, Catherine Huang https://t.co/WqjZfbMv8s 2021-07-27 17:13:59 This works very well. SMAC outperforms prior offline meta-RL methods by a huge margin, and more or less matches the performance of standard online meta-RL (which are not limited in how many episodes w/ rewards they get), even on relatively difficult suboptimal datasets. https://t.co/UHO8R3nvdH 2021-07-27 17:12:26 This is much easier: it requires online data collection, but doesn't require real tasks, hence the online phase is self-supervised. The setup: get *offline* data labeled with rewards, then some self-supervised online data, and then we're done. https://t.co/iRLOBaNHnS 2021-07-27 17:11:46 Fortunately, fixing this problem just requires "reconciling" the adaptation (i.e., learned alg) with the exploration (i.e., the trajs meta-trained policy produces) -- it does *not* require real tasks! So SMAC just invents fake tasks, rolls them out, and uses this to "reconcile" 2021-07-27 17:10:31 The reason is subtle: meta-RL learns to adapt *and* explore, but when it explores for a new task, those exploration trajs look different from the offline data. So any meta-learned alg that is used to adapting to the offline trajs will get messed up when exposed to the online ones https://t.co/Omh5pu9lrK 2021-07-27 17:08:51 Offline RL suffers from distribution shift &lt 2021-07-27 17:07:37 Offline RL+meta-learning is a great combo: take data from prior tasks, use it to meta-train a good RL method, then quickly adapt to new tasks. But it's really hard. With SMAC, we use online self-supervised finetuning to make offline RL work: https://t.co/mtgZTivlVs A thread: https://t.co/mx7caQSrlx 2021-07-25 22:30:22 Website: https://t.co/gMEncxzxzA Paper: https://t.co/URRPDUl6tL Talk video: https://t.co/EbHk7GpHMQ (the paper is currently held up for moderation on arxiv, we'll post the arxiv link as soon as that goes through... its been stuck for several weeks) 2021-07-25 22:29:13 While this sort of "pragmatic compression" is in its infancy, I think it could be tremendously valuable in the future: it's easy to do the A/B testing necessary to train the discriminator, and PICO does not need knowledge of the downstream task, just which action users took. 2021-07-25 22:28:17 This is obviously a proof of concept, and we crank up the compression factor way too high. See below for example: if the user's downstream goal is to check if a person is wearing glasses, eventually PICO scrambles *everything* else, including their gender, but keep the glasses! https://t.co/k3i6TCDHRV 2021-07-25 22:27:03 The compression itself is done with a generative latent variable model (we use styleGAN, but VAEs would work great too, as well as flows). PICO basically decides to throw out those bits that it determines (via its GAN loss) won't change the user's downstream decision. 2021-07-25 22:25:50 The idea is pretty simple: we use a GAN-style loss to classify whether the user would have taken the same downstream action upon seeing the compressed image or not. Action could mean button press when playing a video game, or a click/decision for a website. https://t.co/ldVqu4yf8o 2021-07-25 22:24:55 An "RL" take on compression: "super-lossy" compression that changes the image, but preserves its downstream effect (i.e., the user should take the same action seeing the "compressed" image as when they saw original) https://t.co/gMEncxzxzA w @sidgreddy &amp &gt 2021-07-25 22:21:42 RT @sidgreddy: PICO is a “pragmatic” image compression system that preserves *downstream user behavior* instead of appearance. Idea is to c… 2021-07-24 20:55:59 An extended version of my talk on self-supervised offline RL from today's ws. Discusses how offline RL could provide a general self-supervised pretraining objective. Covers CQL https://t.co/TYL7RTNbO7 AM https://t.co/aLDnuH9an7 VAL https://t.co/Gf5LBr1ZV2 https://t.co/UABGbZOmDg 2021-07-24 16:47:35 I’ll be giving a talk about how reinforcement learning might provide an effective self supervised learning approach at the Self-Supervised Learning Workshop at #ICML2021 at 1:20 pm PT today (Sat): https://t.co/VYuxTXLYj3 Hope to see you there! 2021-07-24 01:54:58 @abhishekunique7 @uwcse @berkeley_ai @pabbeel Congratulations Abhishek! It was great to work together these past years, looking forward to many more collaborations in the future! 2021-07-23 21:32:44 We'll also present conservative data sharing (CDS), a new algorithm that provides a principled way to select which data to share between tasks in offline RL to minimize distributional shift! CDS will also be presented 6 pm PT Sat in the RL theory WS: https://t.co/4mrKiQVQ3r https://t.co/HePVTEVFmR 2021-07-23 21:31:33 We'll present CoMPS, an algorithm for online continual meta-learning, where an agent meta-learns tasks one by one, with each task accelerating future tasks. By @GlenBerseth, WilliamZhang365, @chelseabfinn https://t.co/Fo9an7NJW7 2021-07-23 21:30:11 We'll show how RL can control robots that learn to clean up a room, entirely in the real world. By Charles Sun, @ColinearDevin, @abhishekunique7, @jendk3r, @GlenBerseth. https://t.co/sKjUlGnuhd 2021-07-23 21:29:11 RAIL will be presenting a number of exciting late breaking poster results at the RL4RealLife WS #ICML2021 (8 pm PT today!): https://t.co/dgpL57T6ll Algorithms for real-world RL w/ mobile manipulators, lifelong meta-learning methods, principled multi-task data sharing. A thread: https://t.co/v5r4s5oCHu 2021-07-23 21:24:53 @simsamsom @michaeljanner @qiyang_li You don't have to outrun the mountain lion, you just have to outrun the humanoid that is not using the trajectory transformer 2021-07-23 17:08:17 We'll be presenting how robotic hands can learn complex tasks entirely in the real world at #ICML2021 RL 4 Real Life WS: https://t.co/dgpL57T6ll Fri 7/23 (today), 4pm PT spotlights, 8pm PT poster Full talk here: https://t.co/82u7QxaK3v Paper: https://t.co/iUBV1Cid7K https://t.co/OFEB8f5GdE 2021-07-23 05:05:43 Want to learn about trajectory transformers? @michaeljanner &amp Unsupervised RL 9:30am PT poster https://t.co/70bUUxvF6i A6 RL for Real Life WS 8pm PT poster https://t.co/2OAZWg9XP2 A1 https://t.co/jiMFsZ5sYK 2021-07-23 05:02:19 RT @agarwl_: A thread by @svlevine about this work: https://t.co/MylcoFbTIR Joint work with @aviral_kumar2, @AaronCourville, @tengyuma and… 2021-07-23 03:43:09 @ramav_matsci My guess is probably not, because limited expressivity causes some major issues (see eg https://t.co/utlkdqTcLn). In some sense our result shows that this issue causes big networks to lose expressive power when used in TD updates. 2021-07-23 03:21:42 The workshop paper is here: https://t.co/VpF7tvbXcJ We'll be releasing a full-length paper on arxiv after a few finishing touches and revisions after the workshop. 2021-07-23 03:19:26 This helps across the board for offline RL, for different offline RL algorithms, improves the stability of these methods (making it easier to pick the # of gradient steps), and mitigates the implicit under-parameterization effects that we analyzed in our prior work. https://t.co/xMvpLWmO7F 2021-07-23 03:18:19 This is actually very bad, because time-correlated features alias good actions to bad actions, and lead to horrible solutions. Fortunately, once we recognize this, we can "undo" the bad part of implicit regularization with good *explicit* regularization, which we call DR3. 2021-07-23 03:17:38 We provide both empirical and theoretical analysis that suggests the possibility that the face implicit regularization that makes supervised learning work actually *harms* RL with TD backups, due to bootstrap updates producing time-correlated features. 2021-07-23 03:16:54 One might surmise the same will be true in deep RL: deep RL will work well because it enjoys the same implicit regularization as supervised learning, whatever that might be -- seems reasonable, right? But deep RL seems really finnicky, unstable, and often diverges... 2021-07-23 03:16:07 Deep networks are overparameterized, meaning there are many parameter vectors that fit the training set. So why does it not overfit? While there are many possibilities, they all revolve around some kind of "implicit regularization" that leads to solutions that generalize well. 2021-07-23 03:14:50 You can watch the talk in advance here: https://t.co/NItsnNcI9M And then come discuss the work with Aviral at the poster sessions! This work is not released yet, but it will be out shortly. We're quite excited about this result, and I'll try to explain why. 2021-07-23 03:13:44 In RL, "implicit regularization" that helps deep learning find good solutions can actually lead to huge instability. See @aviral_kumar2 talk on DR3: 7/23 4pm PT RL for real: https://t.co/4mrKiQVQ3r 7/24 5:45pm PT Overparameterization WS talk https://t.co/1lN1IIwbVs #ICML2021 &gt 2021-07-22 00:02:32 @CsabaSzepesvari @lawrennd @tdietterich @rodneyabrooks Worth mentioning that the specific strategy matters: they focused on a *very* simulation heavy approach, which doesn't benefit from real-world data and does not have any notion of "flywheel" -- unsurprising that this would be incompatible with the rest of what they do. 2021-07-22 00:00:29 @CsabaSzepesvari @lawrennd @tdietterich @rodneyabrooks I'm also uneasy about this take (though perhaps there is a phrasing ambiguity). I think there has been lots of progress on motor ctrl and robotics since 2015, including in RL. OpenAI picked a strategy, which didn't pan out for them, that doesn't mean no progress is happening. 2021-07-21 18:45:55 (small typo, "algorithmic causality" in the first post should have been "algorithmic independence," which is a property used to *show* causality) 2021-07-21 18:44:53 When actions compete over who gets to go, the right type of auction will cause them to bid their real Q-value! And this preserves modularity, leading to better transfer. You can also read about our previous work on CVS here: https://t.co/7tTgAsChV0 2021-07-21 18:44:13 It turns out that societal decision making (CVS) from our previous paper has this property. In CVS, each action is a different "agent," and the actions "compete" to decide who goes: each action bids some utility, the winner goes and collects reward (positive or negative). https://t.co/Gjlfr16tEy 2021-07-21 18:43:11 Q-learning is better than policy gradient, and tabular Q-learning *is* modular (because values at different states are independent), but with function approximation, we need some kind of independence between the mechanisms that choose different actions. 2021-07-21 18:42:07 The basic idea is very simple: if decisions are made by independent mechanisms, then transfer is easier if only a few of those mechanisms need to change. What is interesting is which existing algorithms have this kind of independent mechanisms property. 2021-07-21 18:41:23 Want to learn how algorithmic causality leads modular RL methods to transfer better? Check out @mmmbchang's long talk at #ICML2021 tomorrow 7/22 at 5:00 am PT, or watch it right here: https://t.co/fz69J6KMWg https://t.co/aNM5T37e1E https://t.co/lmyhZTgBs3 Short thread: https://t.co/Aqdh1hjktU https://t.co/N4dO1baLOi 2021-07-21 04:11:57 Meta-learning can do some pretty useful things when applied properly :) Congrats to @mike_h_wu, @chrispiech &amp 2021-07-20 19:31:16 This work is philosophically related to some of our prior work on surprise minimization (i.e., maintaining homeostasis) as a way to get intrinsic motivation: https://t.co/H2aL3n7it0 2021-07-20 19:30:24 Adversarial surprise: an "intrinsic motivation arms race" where Agent A tries to surprise Agent B, and Agent B wants to avoid getting surprised. Both a way to get emergent complexity, and provide some appealing coverage guarantees for exploration. Check out the paper below: https://t.co/USsuRlmvGH 2021-07-20 16:41:16 Model-based RL with images and actual trajectory optimization methods, that are more sophisticated than random whatever (sorry, CEM craze is partly my fault... but we really can do better!). Cool thing here is that collocation methods can be applied directly to latent states. https://t.co/aEuj03ivAn 2021-07-20 16:37:54 This work on learning by observing other agents presented at #ICML2021, 5:25 pm PT today (7/20) https://t.co/cMf5ZASnuC Can RL exploration &amp 2021-07-20 01:05:17 Spotlight is Tue Jul 20 05:35pm PT here: https://t.co/rm4imSshjt Arxiv: https://t.co/aLDnuH9an7 Video: https://t.co/7z1x9ppmoH My favorite thing about this paper is how it gives a glimpse into how goal-conditioned RL can provide "universal offline self-supervised pretraining" 2021-07-20 01:03:35 Goal-conditioned offline RL, with real data, can enable robots to perform diverse tasks and provide general self-supervised objectives for RL that help with dwnstream tasks. See @YevgenChebotar's talk on this (Actionable Models) Tue 7/20 5:35pm #ICML2021: https://t.co/GMsIUMZlPM https://t.co/wyp8YbZ7fW 2021-07-19 23:49:19 RT @shiorisagawa: Just in time for ICML, we’re announcing WILDS v1.2! We've updated our paper and added two new datasets with real-world di… 2021-07-19 19:00:06 This makes calibration much better on OOD inputs, and scales well even to very large conv nets! You can find out more in Aurick's blog post: https://t.co/ES4l6oP6vJ The ICML spotlight (6:25 am PT Tue 7/20) is here: https://t.co/zv2uUAs9Ph 2021-07-19 18:58:36 This is not new: NML is a well known concept. In our paper, we show how to make it tractable for deep nets. Instead of retraining the model for each test point, we pretrain a Bayesian posterior over parameters, and finetune on *only* the new test point with this regularizer https://t.co/wcCygKwQfX 2021-07-19 18:57:16 If we visualize the normalized max likelihood (ratio of probabilities of models assigning different labels), we see that points far from the data are highly uncertain (unlike the original model, which is very certain but very incorrect). Here shown for different reg levels lambda https://t.co/P11NAbmXBF 2021-07-19 18:55:43 It's easiest to get an intuition for this in a 2D logistic regression example. When we label the pink point (lower right) with either class, the decision boundary changes a lot -- the model class can explain *either* label with high confidence, so the label is unknown. https://t.co/wyGFHT9QFA 2021-07-19 18:54:43 Let's say we have an OOD input (below), and we don't know what class it is. Standard neural nets might give an overconfident response. The idea in (C)NML is to label it with each label, finetune the model, and then use the ratio of resulting likelihoods. https://t.co/K0bt3qdDWW 2021-07-19 18:53:28 Training on the test set is usually bad, but with amortized conditional normalized max likelihood (ACNML), it can improve calibration on OOD inputs (without "cheating") -- Aurick Zhou will present ACNML at #ICML2021 tmrw Tue 7/20 6:25 am PT https://t.co/piwcU5wMs4 Short summary: 2021-07-19 17:36:09 RT @brandontrabucco: Exciting collaboration with @aviral_kumar2, @YoungGeng, and @svlevine in BAIR! Want to design drug molecules, protein… 2021-07-18 19:25:48 @EliSennesh Kevin Li and @abhishekunique7 made them, my contribution was to annoy the heck out of them by telling them to make the figures better (with minimal info how to do so) There are definitely better tools, but #1 advice (as with any art/presentation thing) is to iterate a lot 2021-07-18 19:21:33 We'll present MURAL at #ICML2021. Tue Jul 20 07:35 PM -- 07:40 PM (PDT): https://t.co/VOj1A2PtVz w/ Kevin Li, @abhishekunique, Ashwin Reddy, Vitchyr Pong, Aurick Zhou, Justin Yu 2021-07-18 19:19:50 This ends up working very well across a wide range of manipulation, dexterous hand, and navigation tasks. To learn more about NML in deep learning, you can also check out Aurick Zhou's excellent blog post on this topic here: https://t.co/ES4l6oP6vJ https://t.co/gum3xgd5R7 2021-07-18 19:18:37 Doing this tractably is hard, because we need two new classifiers for *every* state the agent visits, so to make this efficient, we use meta-learning (MAML) to meta-train one classifier to adapt to every label for every state very quickly, which we call meta-NML. https://t.co/ehBLef43QB 2021-07-18 19:17:50 This provides for exploration, since novel states will have higher uncertainty (hence reward closer to 50/50), while still shaping the reward to be larger closer to the example success states. This turns out to be a great way to do "directed" exploration. 2021-07-18 19:16:35 This is where the key idea in MURAL comes in: use normalized max likelihood (NML) to train a classifier that is aware of uncertainty. Label each state as either positive (success) or negative (failure), and use the ratio of likelihoods from these classifiers as reward! https://t.co/CGKTddhpEJ 2021-07-18 19:14:59 The website has a summary: https://t.co/wE2ybwVYBn If the agent gets some examples of high reward states, we can train a classifier to automatically provide shaped rewards (this is similar to methods like VICE). A standard classifier is not necessarily well shaped. https://t.co/kX5ULbXiUy 2021-07-18 19:13:25 Can we devise a more tractable RL problem if we give the agent examples of successful outcomes (states, not demos)? In MURAL, we show that uncertainty-aware classifiers trained with (meta) NML make RL much easier. At #ICML2021 https://t.co/dsD7nLKggG A (short) thread: https://t.co/KzqXZpvt7D 2021-07-18 02:04:54 NEMO was published in ICLR 2021, and you can find the ICLR 2021 presentation here: https://t.co/F4kMV2fZai If you want to learn more about NML and how it can be used in supervised learning, also check out Aurick Zhou's blog post: https://t.co/ES4l6oP6vJ 2021-07-18 02:04:41 This corresponds to the normalized maximum likelihood (NML) distribution, which has appealing regret guarantees, which we extend in NEMO to provide regret guarantees on offline MBO as well! This is more complex than COMs, but potentially more powerful as we get a full posterior. 2021-07-18 02:04:20 The basic idea, unlike COMs (which learn pessimistic models) is to get a posterior over values for a new design x. Justin's method (NEMO) trains a separate model *for every possible value y* for the design x (discretized), and uses the likelihood from these to get the posterior. https://t.co/qvY7ps0cO5 2021-07-18 01:57:57 Since many people were interested in our recent offline MBO work, I'll also write about a recent paper on MBO by Justin Fu, which trains forward models for each possible objective value and uses them to compute a posterior via NML: https://t.co/I9I8SRNLXs A thread: 2021-07-18 01:54:36 @bodonoghue85 We didn't try generating images with this approach, that could be interesting to try -- a number of recent approaches to image generation with EBMs employ similar style training objectives. We did study image generation as MBO in our prior work MINs: https://t.co/z3la4MGvUw 2021-07-16 21:18:47 @alfcnz If you have any ideas, let us know :) 2021-07-16 21:18:23 @alfcnz Yes, absolutely. The same energy model connection also holds for conservative Q-learning (CQL): https://t.co/TYL7RTNbO7 The form of the gradient is the EBM gradient. We're not sure what to do with this connection, and our theory doesn't really use it, but it is intriguing. 2021-07-16 21:01:18 We'll be presenting COMs at #ICML2021: Thu Jul 22, 6:40am-6:45am PT https://t.co/Bgff1gLseh Poster: Thu Jul 22, 6am - 8 am PT Or just watch the talk in advance here: https://t.co/JxIMy2hlaB 2021-07-16 20:59:54 Great collaboration w/ Brandon Trabucco, @aviral_kumar2, @YoungGeng. For more offline MBO papers, also check out these from our lab: https://t.co/z3la4MGvUw https://t.co/I9I8SRNLXs And these great papers from @jlistgarten's group: https://t.co/5oVYjknwD1 https://t.co/ZsEK3TC5jR 2021-07-16 20:56:44 Just like CQL, we can guarantee that for a suitable regularizer weight, COMs gets us a lower bound on the true value of a design, providing a principled solution. It's also simple to use and implement. Website: https://t.co/v14SkeEIHD ICML talk: https://t.co/JxIMy2hlaB 2021-07-16 20:55:41 Our method, conservative objective models (COMs), is very similar to adversarial training: just find "adversarial" designs that the model thinks are good, and make them look bad, while making the designs in the data look good to counterbalance this. https://t.co/jjiU3hVtYG 2021-07-16 20:54:48 Just like in offline RL, the problem is distributional shift: the new designs we want to evaluate might be out of distribution for our training data. So we can borrow some ideas from conservative Q-learning (CQL) to "robustify" our models, thus making them suitable for design. 2021-07-16 20:54:07 This is very important: lots of recent work shows how to train really good predictive models in biology, chemistry, etc. (e.g., AlphaFold), but using these for design runs into this adversarial example problem. This is actually very similar to problems we see in offline RL! 2021-07-16 20:53:15 The basic setup: say you have prior experimental data D={(x,y)} (e.g., drugs you've tested). How to use it to get the best drug? Well, you could train a neural net f(x) = y, then pick the best x. This is a *very* bad idea, because you'll just get an adversarial example! 2021-07-16 20:52:17 Data-driven design is a lot like offline RL. Want to design a drug molecule, protein, or robot? Offline model-based optimization (MBO) tackles this, and our new algorithm, conservative objective models (COMs) provides a simple approach: https://t.co/ivfCvN147Z A thread: https://t.co/1olEYCE74y 2021-07-16 03:19:52 Check out sigma-VAE at #ICML2021 next week, for some simple and effective methods for training calibrated VAE decoders. Poster: Thu 22 July 17:35 - 17:40 PT https://t.co/L3rJFPWttI Paper: https://t.co/4ED8dsKr9A Web: https://t.co/Dx55xb1Jv3 Talk: https://t.co/5WRvaPbnae https://t.co/YtCeX7C7WP 2021-07-15 17:48:28 Bringing together IRL and multi-agent: look at other agents &amp https://t.co/2aCBKW2Rng https://t.co/5XcPN19BkB https://t.co/DEoyfKT9TC 2021-07-14 18:19:03 @qcjefftqc That is a really good question! I do think it might provide part of the explanation, though such methods do not (in principle) capture the right "type" of uncertainty. But it seems likely they "accidentally" get some epistemic uncertainty, esp. with regularization/early stopping. 2021-07-14 18:03:23 @gershbrain @its_dibya Yup, fair enough, thanks! I didn't actually know about the '60s Martin &amp 2021-07-14 17:58:53 @gershbrain @its_dibya Oh, I see -- the POMDP formulation of Bayesian RL is indeed well known (this is in the paper too). We did not see prior work that studies this specifically in the context of generalization (vs explore/exploit tradeoffs), though if you know of any, def. that should be referenced. 2021-07-14 17:44:59 This was a really fun collaboration with @its_dibya, @jrahme0, @aviral_kumar2, @yayitsamyzhang, @ryan_p_adams -- a really fun group to work with on generalization 2021-07-14 17:43:32 Unfortunately, solving (or even estimating) the epistemic POMDP is very hard, and LEEP makes some very crude approximations. Lots more research is needed to utilize the epistemic POMDP, in which case I think we can all make lots of progress on generalization! 2021-07-14 17:42:34 Based on this idea, we developed a new algorithm, LEEP, that utilizes epistemic POMDP ideas to get better generalization. LEEP actually does *worse* on the training environments, but much better on test environments, as we would expect. https://t.co/l5AxpWI1AQ 2021-07-14 17:41:00 What is happening here is that generalization in RL requires taking epistemic (information-gathering) actions at test time, just like we would in a POMDP, but this is never optimal to do at training-time. Hence, MDP methods will not generalize as well as POMDP methods. 2021-07-14 17:40:16 This leads to some counterintuitive things. Look at the "zookeeper example" below: the optimal MDP strategy is to look at the map (which is fully observed) and go to the otters, but peeking through windows generalizes much better (is never optimal in training). https://t.co/2lLv65hTee 2021-07-14 17:38:52 The guessing game is a MDP, but learning to guess from finite data becomes (implicitly) a POMDP -- what we call the epistemic POMDP, because it emerges from epistemic uncertainty. This is not unique to guessing, the same holds eg for mazes in ProcGen, robotic grasping, etc. 2021-07-14 17:38:08 Of course, this is a bad idea -- if it guesses wrong on the first try, it should not guess the same label again. But this task *is* fully observed -- there is a unique mapping from image pixels to labels, the problem is that we just don't know what it is from training data! 2021-07-14 17:37:25 Take a look at this example: the agent has a multi-step "guessing game" to label an image (not a bandit -- you get multiple guesses until you get it right!). We know in MDPs there is an optimal deterministic policy, so RL will learn a deterministic policy here. https://t.co/zih0Jp5h6l 2021-07-14 17:36:33 Empirical studies observed that generalization in RL is hard. Why? In a new paper, we provide a partial answer: generalization in RL induces partial observability, even for fully observed MDPs! This makes standard RL methods suboptimal. https://t.co/u5HaeJPQ29 A thread: 2021-07-14 17:36:29 RT @its_dibya: Fresh on ArXiv! https://t.co/KY9tnogl6Z TL 2021-07-14 17:32:37 @gershbrain @its_dibya Bayes-adaptive POMDPs are formulated explicitly as POMDPs. The epistemic POMDP emerges from solving a fully observed MDP with finite samples. This requires new terminology: the partial observability emerges from epistemic uncertainty, even though the MDP is not partially observed 2021-07-12 21:27:50 Also, it really hammers home the notion that I might be living in some sort of science fiction fantasy when our work literally involves enabling robots to play with stuff. Though having to write a grant proposal tonight somewhat interferes with the illusion. 2021-07-12 21:15:23 The ideal robotics lab is one where there are lots of robots playing, and a few humans working (with the largest ratio of the former to the latter) 2021-07-12 21:13:13 For those of you wondering what project this is, it's the next stage of our reset-free dexterous manipulation work -- the first phase publication can be found here: https://t.co/iUBV1Cid7K 2021-07-12 21:12:30 Always makes my day to come in to lab and see a robot playing with an object in the corner, without adult supervision Though I think this one needs a friend, it's been going at it for some time on its own... https://t.co/cA9bCHFQE8 2021-06-30 17:02:15 To learn more about HInt, check out the: website: https://t.co/srPyLD7jql paper: https://t.co/HLTTjJD11S video: https://t.co/a525gRXKrG 2021-06-30 17:02:15 To learn more about HInt, check out the: website: https://t.co/srPyLD7jql paper: https://t.co/HLTTjJD11S video: https://t.co/a525gRXKrG 2021-06-30 17:01:31 In experiments, we can combine data from three different robots (a TurtleBot, Clearpath Jackal, and a drone), as well as data collected by a person walking with a camera! This helps each robot generalize, even to environments that it hasn't seen before (but others have). https://t.co/2XMTkTJgVB 2021-06-30 17:00:24 This is a model-based method, but it doesn't predict future observations, rather future positions and rewards (similarly to GCG, CAPS, BADGR, and DIMs): https://t.co/QjLiykn3gr https://t.co/qoxAa0oIaK https://t.co/1iQmbzqWPT https://t.co/XGiEcvA78h 2021-06-30 16:58:29 Then, we can compose a robot-specific dynamics model with the shared visual model at test-time to get a single model that predicts future outcomes (e.g., rewards) given the robot's specific motor commands, and still share the visual part across many robots! 2021-06-30 16:57:55 The basic idea is that we learn two models: a model that maps images &amp 2021-06-30 16:56:31 Can we combine data from many robots and learn a single model-based controller? In HInt (Hierarchically Integrated models), Katie Kang &amp https://t.co/HLTTjJD11S -&gt 2021-06-30 04:23:12 The intuition behind the result is that, if each virtual "agent" has a different "job," and these jobs are independent, then the whole society can adapt more quickly when transfer requires changing just a sparse subset of these jobs. 2021-06-30 04:22:24 In this new paper, we further show that such decentralized and modular reinforcement learners satisfy an algorithmic independence property that allows them to adapt much more quickly in transfer learning scenarios, by utilizing tools for studying independent causal mechanism. 2021-06-30 04:21:20 In our prior work, we showed that we can design populations ("societies") of virtual agents that compete to get the largest share of extrinsic reward, "bidding" over which action to take at each time step. We called this the "cloned Vickrey society": https://t.co/PKOBuatwYK https://t.co/6Jf4nRK2i5 2021-06-30 04:19:01 Algorithmic independence can explain why modularity can result in better transfer -- in this paper, we study this for the particular case of modular algorithms, building on the our prior work on a "societal competition" model of RL (which is inherently modular). -&gt 2021-06-25 05:58:27 Frustratingly simple video prediction first time I saw overfitting as the main problem for video prediction (though data augmentation can address that). Great results on video pred many frames into the future, and also works for real-world robot control! (and yes there is code) https://t.co/jxulSzweSG 2021-06-25 04:45:07 Visual model-based RL with a real trajectory optimizer -- collocation + visual model-based RL can "imagine" moving objects to the goal, and then figure out how to actually accomplish it. See below for @_oleh's excellent summary, check out the paper here: https://t.co/9AGIzCY9jz https://t.co/Q7vecze7K3 2021-06-22 04:06:55 This is perhaps a bit sad, because I(Z[t+1] 2021-06-22 04:04:54 What *is* sufficient? Turns out that under some mild assumptions, I(Z[t+1] 2021-06-22 04:03:45 Another common choice is MI between past states and future states: I(Z[t+1] 2021-06-22 04:02:00 A popular choice is MI between between actions and future states: I(A[t] 2021-06-22 03:59:39 Mutual info maximization is often used for representation learning. Which MI objectives lead to representations that are sufficient to learn optimal policies in RL? In recent work by Kate Rakelly et al we show some popular choices are not sufficient: https://t.co/Wg1vUe8Ouh &gt 2021-06-17 03:17:47 The updated meta-world has significantly improved visuals and many bug fixes, improved reward functions, and other changes to make it easier to work with. If you're looking for 50 robotic manipulation tasks for meta-learning &amp 2021-06-16 22:42:11 A kind of GAN for system ID: Train simulator parameters so that the simulated *trajectories* are indistinguishable from real trajectories, treating the simulator as a "policy" that tries to replicate long-horizon behavior seen in the real data. https://t.co/6LcPukaVY7 https://t.co/F0BHJspW8u 2021-06-15 03:15:33 Prior to this, I didn't think that goal-conditioned RL could handle such diverse real-world scenes. But what's neat here is not just the vis. goal results, but that it provides pretraining &amp 2021-06-14 20:28:11 RT @minerl_official: We are so excited to announce that the first round of the 2021 @NeurIPSConf MineRL Diamond Competition has officially… 2021-06-09 03:24:14 What I find most interesting about this work is the notion that information-theoretic skill learning (e.g., DIAYN) can be thought of as a kind of goal-conditioned RL with a learned representation. This opens up a number of interesting design choices. https://t.co/bdhzbMw0kM 2021-06-08 19:13:27 RT @siggraph: This #SIGGRAPH2021 Technical Papers selection will really get you moving with new #research on character control in "AMP: Adv… 2021-06-04 20:31:00 @rsalakhu That is a really beautifully illustrated blog post! ...is that a pig walking across the field in the Unreal Engine/AirSim screenshot? 2021-06-04 18:00:04 @FeryalMP @AlexKhazatsky @ashvinair @khimya Thank you for pointing out, definitely an oversight on our part that we should rectify. But aside from the title (which I agree is an unfortunate collision, though unintentional on our part), the two papers are trying to do very different things. Which is why we missed it. 2021-06-04 06:56:52 I'll give a talk tomorrow at ICRA Safe Robot Control workshop https://t.co/puXIQyDVHj (Zoom link on website) 9:30 am PT Fri 6/4! "Safety in Numbers: How Large Prior Datasets Can Put RL into the Real World" (it's a workshop talk, so this title is perhaps a bit aspirational). 2021-06-03 20:49:40 @AlbertBuchard I suspect it depends a lot on the data, but the analysis in Figure 2 suggests that at least for prediction accuracy, it's about having better state representations rather than being non-Markovian (the story may be different if we look at actions, depending on the behavior policy) 2021-06-03 20:32:40 @ethancaballero The geographic extent of BAIR while everyone is working from home (in various locations) has a radius of roughly 12,000 miles. 2021-06-03 20:22:41 4/ Concurrently (just yesterday!) Decision Transformer validated the same hypothesis: that big sequence models can do offline RL effectively. There are some differences (roughly, trajectory transformer is more model-based), but the conclusion is the same: https://t.co/cnfyxh85ZE 2021-06-03 20:22:00 3/ Interestingly, the model is significantly better than prior models in model-based RL for forward prediction too, able to predict humanoid trajectories with physical accuracy for 100 steps into the future (typically, model-based RL rollouts are much shorter than this) https://t.co/ZqZr83uRoh 2021-06-03 20:20:49 2/ For control, we can just decode with beam search, but using reward instead of likelihood, which is competitive with or better than prior offline RL methods on standard benchmarks, can solve goal-conditioned tasks, etc. https://t.co/t0dgl8uf6O 2021-06-03 20:20:29 1/ The basic idea is simple: take all dimensions of states, actions, and rewards, discretize them, and model them as a long sequence of discrete tokens. Then use a language model. This can be used to predict, and also to control. https://t.co/5RKQrOgitf 2021-06-03 20:20:10 Can we replace RL with one big sequence model? Trajectory transformer models state/action/reward sequences one token (dimension) at a time -- this allows long-horizon prediction and offline RL w/o separate actors, critics, constraints, etc.: https://t.co/98D21UVSMC A thread: https://t.co/YwJeVL02tZ 2021-06-03 05:48:28 This will be in the session "Reinforcement Learning for Robotics I" at 19:30 GMT+1/11:30 am PT Thu 6/3. You can find the paper here: https://t.co/SDEtxYcxgd Code: https://t.co/uPVACeN9cI Video: https://t.co/I1iyXQYSXt 2021-06-03 05:47:24 We do a lot of work on manipulation. But what about aerial manipulation? Can we use meta-learning to get drones to pick up cargo? At #ICRA2021 19:30 GMT+1/11:30 am PT Thu 6/3, Suneel will present meta-learning w/ suspended payloads: https://t.co/0DgaNGWgSJ https://t.co/uQdhgZJHog 2021-06-03 04:18:46 From driving on sidewalks and paved paths, we can also learn to reach visually indicated goals with a mobile ground robot. Check out @_prieuredesion's ViNG talk, #ICRA2021 tmrw 1900 GMT+1/11am PT, "Autonomous Vehicle Navigation II" https://t.co/XPrUpNSHhz https://t.co/hyTUbxavG1 https://t.co/Ib0E6IXIph 2021-06-03 04:15:55 RT @ashvinair: Super excited to release latest work on visuomotor affordance learning (VAL)!! Highlights: 1. VAL utilizes a large prior dat… 2021-06-03 03:26:53 Check @nick_rhinehart's talk at ICRA tmrw 18:00 GMT+1/10 am PT (Thu 6/3)! Find out how robots can plan for contingencies, handling autonomous driving scenarios where humans can react in unpredictable ways. If you can't make the talk, recording is here: https://t.co/ol3WmIziGy https://t.co/xhAdLSaR2E 2021-06-02 03:23:32 Paper: https://t.co/Gf5LBr1ZV2 Website: https://t.co/pf7qqtM8i1 Video: https://t.co/XthcnLGsii VAL will be presented at #ICRA2021 on Thu, 19:15 GMT+1/11:15 am PT, in the Vision-Based Manipulation session. 2021-06-02 03:21:35 Here are a few animations of new tasks that VAL masters in unseen test environments: for each task, the robot was first allowed to interact without any supervision, using its offline-trained policy + generative model to propose &amp 2021-06-02 03:20:31 We train VAL on a large dataset of diverse environments with many different objects, using offline RL (based on AWAC https://t.co/JYyprRInhR), with a conditional VQ-VAE model for candidate goal generation. After offline training, it autonomously learns in new settings. https://t.co/UWT2xES0Dl 2021-06-02 03:18:38 Our method, visuomotor affordance learning (VAL), trains a generative model that predicts what the robot *could* do, and a goal-conditioned policy to do it. In a new environment, it then imagines what might be doable, and practices doing it. https://t.co/yMJFqN8DvO 2021-06-02 03:17:14 When a robot is put in a new scene, can it "imagine" what tasks could be possible, and then practice them? In "What Can I Do Here?", @AlexKhazatsky, @ashvinair, &amp 2021-06-02 01:04:23 For those who want to hear the whole thing: Spotify: https://t.co/f9XX1btw9c Soundcloud:https://t.co/3kTgd6KEsZ YouTube: https://t.co/Ha1nWCS7M7 2021-06-02 01:03:06 A fun podcast, thank you @MarwaEldiwiny for inviting me! https://t.co/7uDkQtR5cj 2021-06-01 18:56:49 @SmallpixelCar @OpenRoboticsOrg Not yet, we were not brave enough to try to train it to do that -- we only have one of these. There is nothing technical that in principle would prevent this though, just a bit beyond the production and QA capabilities of a university lab. 2021-06-01 05:45:51 At #ICRA2021, @abhishekunique7 will present present our work on dexterous hands, showing that by practicing many skills at the same time, such robots can learn without human intervention! 21:15 GMT+1/1:15 pm PT in Manipulation: Reinforcement Learning II https://t.co/iUBV1Cid7K https://t.co/2XmgmXzDBV 2021-06-01 05:13:29 Also at #ICRA2021 tmrw, S. Nasiriany will present Disco RL, which extends goal-conditioned RL to train policies conditioned on arbitrary goal distributions! 20:15 (GMT+1)/12:15 pm PT Tue, in Manipulation: Reinforcement Learning I https://t.co/LVchQkoIGC https://t.co/KzbaNiFR1V https://t.co/4GXPo17vRT 2021-06-01 04:33:24 Can robots to learn to drive on Berkeley sidewalks, entirely from images? Also at #ICRA2021, Tue 18:30 GMT+1/10:30 am PT, Greg Kahn will present LaND, which can learn to navigate kms of Berkeley sidewalks! In the Field Robotics: Control session. https://t.co/ZujiRF6PGt https://t.co/bmx6ky924G 2021-06-01 04:02:02 Can we learn image-based open-world navigation from diverse data in the real world? At #ICRA2021, Tue 10:45 GMT+1/2:45 am PT, Greg Kahn will present BADGR, our system for real-world navigation, in the Service Robotics Award Session! https://t.co/zHleTYhO1C https://t.co/gdZNWHlkHe 2021-06-01 03:11:01 RT @HelgeRhodin: Will reinforcement learning solve computer vision? With or w/o simulation? @svlevine showcased his recent work and shared… 2021-05-29 17:58:20 RT @siggraph: Quench your #research thirst with an interview with Jason Peng + Edward Ma about "AMP: Adversarial Motion Priors for Stylized… 2021-05-27 07:08:27 @_sam_sinha_ Might have solved AGI, might have trained on the test set, stand by... 2021-05-27 07:01:46 Really stoked to share some of the results in the papers we're submitting on Friday, once we get this nicely written up. We've figured out some things. ... ... ...unless we realize before that that we were wrong, that's always a possibility... 2021-05-26 17:10:17 @F_cossio @abhishekunique7 Each task's policy is only rewarded for that task, not for the performance of future tasks. 2021-05-25 17:40:04 Want to know how robots can learn to give you a hand with your NeurIPS submissions? So do I. In the meantime, you can check out @abhishekunique7's ICRA 2021 talk, how to train robotic hands to do lots of other stufffrom scratch, in the real world https://t.co/vDxDnvNgJr 2021-05-12 17:52:12 Congratulations Prof. Berseth!! For students: working with Glen as he starts his lab at @Mila_Quebec will be a lot of fun. Check out some of his work here: https://t.co/mrSdXe4e39 From control of walking robots to reversing the second law of thermodynamics https://t.co/y3vMyGmJbt 2021-05-12 03:24:29 Of course, these experiments in MT-Opt take that to a whole new level: once the robot is trained on a huge dataset, it actually starts to generalize in meaningful ways, so that it actually reacts *intelligently* when we mess with it (whereas Brett just did funny random stuff). 2021-05-12 03:23:47 What I enjoyed most when I ran robot experiments (before I got too busy to do it myself) was messing with a robot that had learned a task. It's so much fun to see what it tries to do in different situations. E.g., these demos (w/ @chelseabfinn &amp 2021-05-08 18:46:40 RT @hausman_k: Actionable Models got accepted to #ICML2021! Excited to talk about this work at the conference! https://t.co/I1Mg40AupR 2021-05-07 22:02:00 @davidgordian I suspect it could work. Karl Zipser at Berkeley had a cheap little RC car that he was using for a while for research along similar lines, which seemed to work pretty well -- I'm not sure if he ever open-sourced the design, but I think it was a few hundred $ in cost. 2021-05-07 19:25:06 This is starting in 36 min (1 pm PT) here: https://t.co/It721zZYoM Meanwhile, check out Abhinav Gupta's talk, which starts shortly (12:30 pm PT)! #ICLR2021 https://t.co/cwDQ92ShP5 2021-05-07 05:43:36 @kolari My recommendation to the students to tackle this particular problem for the evaluation had absolutely nothing to do with the fact that I have a toddler at home. Completely. not. related. 2021-05-07 04:56:50 I'll be giving a talk on how robots can learn autonomously from diverse datasets tomorrow (Fri) at 1 pm PT at the Embodied Multimodal Learning workshop! There will be many robots, learning many things. ICLR page: https://t.co/It721zZYoM https://t.co/JFOt1vtJXD 2021-05-07 03:29:15 @AleksandraFaust @iandanforth @JDCoReyes Oh wow good catch, I didn't even know he had one 2021-05-07 02:36:08 In "RL for Autonomous Mobile Manipulation with Applications to Room Cleaning" we show how a hierarchical RL approach can enable continual, unattended real-world learning: https://t.co/20fxjDtfD6 2021-05-07 02:34:57 Fri 6 am PT &amp https://t.co/VkNGqn73gP https://t.co/51Az37MsXx 2021-05-07 01:51:14 @iandanforth He's too cool for us (or maybe just too busy getting his papers invited to workshops and getting ICLR oral presentations... while I'm just here posting on twitter) 2021-05-07 01:22:03 To improve RL, we try to build better algorithms. But what if we ask how to set up environments to make RL (esp. lifelong RL w/ sparse rewards) easier? Check out Ecological RL, which was invited to the never-ending RL ws: https://t.co/VkNGqn73gP #ICLR2021 https://t.co/35ibxthqyY https://t.co/b3GBHA35g8 2021-05-06 23:44:32 Robots can learn on their own, if they reset their own task. Multi-task learning is a great way to do this -- check out the poster by @abhishekunique7 et al tmrw 6am PT &amp https://t.co/VkNGqn73gP Web: https://t.co/iUBV1Cid7K Video: https://t.co/IfXEhwAybN https://t.co/sCqpoysa1u 2021-05-06 21:46:53 Can causality and algorithmic independence help RL transfer better? Tmrw, @mmmbchang will present "Modularity in RL via algorithmic independence" in #ICLR2021 ws: Generalization beyond... 1 pm PT: https://t.co/lraDFJO45r Learning to learn 8:40 am PT: https://t.co/T8cNaMFUft https://t.co/qra9AbhJh3 2021-05-06 20:21:52 One of the things that I find particularly exciting about this is that it demonstrates the power of offline RL: we used the same dataset collected by Greg Kahn in spring 2020 to build this navigation system, even though that data was collected for a completely different project. 2021-05-06 20:17:37 How can robots learn to search for goals in new environments? Tmrw, Fri 10:50am PT, @_prieuredesion will present a full-length talk at the NERL ws at #ICLR2021 on RECON, which addresses this problem with real-world data! WS: https://t.co/N6AMQvMlpE paper: https://t.co/Es7YKyv8KX https://t.co/sgHPoT0SmE 2021-05-06 04:50:37 Want to know how to train goal-reaching policies w/o reward, with just state-prediction density? See Ben Eysenbach's poster on C-Learning Thu 9 am PDT at #ICLR2021! https://t.co/fAL5Tea6kk Web: https://t.co/GUhYjmAIXC Code: https://t.co/5tYQKx5kGw ICLR: https://t.co/6DbcScG7bL https://t.co/mVSSqwveSb 2021-05-05 17:37:09 How can RL agents explore *safely*? Conservative safety critics aim to provide this via a pessimistic safety critic that believes that anything you haven't done before is dangerous. At #ICLR2021 today, 5 pm PT, rm B5! arxiv: https://t.co/qsYhXBeipZ ICLR: https://t.co/2Oqg8mDF20 https://t.co/ejTBMHCkT7 2021-05-05 17:33:06 Today (Wed) at 11:45 am PT, JD Co-Reyes et al. will present Evolving RL Algorithms -- how we can search over the space of Q-learning methods to discover more effective objective functions! Poster at 5 pm PT https://t.co/pmMBI1gAmz https://t.co/3MzyubeqdP https://t.co/5RAOHC6YNt https://t.co/Mo8mgT5XcP 2021-05-05 03:49:58 And by "Tue 9 am PT" I actually mean "Wed 9 am PT", oops 2021-05-05 03:07:47 Can hierarchical latent variable models help solve offline RL problems? Tue at 9 am PT, Anurag Ajay et al. will present OPAL, a method that combines learned latent action models with higher-level offline RL. Paper: https://t.co/8phIcOMkyC Poster: https://t.co/zA4oGud51H https://t.co/tYekSrTL8S 2021-05-04 21:15:58 Deep RL has an _implicit_ under-parameterization problem: TD backups cause aliasing in the features, even for large nets! Today 5 pm PT, @aviral_kumar2 and @agarwl_ present analysis of IUP and some solutions! #ICLR2021 ICLR: https://t.co/2sg6WRLjT2 web: https://t.co/NogUM92z1o https://t.co/kvoj6um3C4 2021-05-04 19:54:32 Interested in offline model-based RL with images and goals? Check out MBOLD, presented at #ICLR2021 this evening (Tue 7:15 pm PT): https://t.co/OfHRyV4M9M S. Tian will describe how models + temporal distance values can allow diverse visual goal-reaching. https://t.co/8XuwTI4oaN https://t.co/D7ZNR55mhP 2021-05-04 17:47:02 The great thing about this is that the algorithm is _extremely_ simple. Paper: https://t.co/fn8sEiLOsT Blog: https://t.co/8wZp0pEiOx Poster: https://t.co/Jf0gqpsMu3 Talk recording (if you want a sneak-peek!): https://t.co/XqRcNINlz5 2021-05-04 17:46:38 The basic idea in GCSL is that we can imitate our own prior experience with suitable relabeling, and this actually corresponds to an RL algorithm that we can analyze, show per-iteration improvement, etc.! And then we can use it to learn all sorts of RL tasks, like lunar lander. https://t.co/UeGnNrPB3l 2021-05-04 17:46:20 Can regular *supervised* learning solve RL problems from scratch (no demos, no prior data)? Does that even make sense? Well, yes! Check out the GCSL presentation at #ICLR2021, poster at 5 pm PT Tue (today), oral presentation 11 am PT Wed: https://t.co/Jf0gqpsMu3 short thread-&gt 2021-05-04 05:08:55 You can also find out more about this (admittedly somewhat peculiar) unsupervised RL method here: Paper: https://t.co/H2aL3n7it0 Project website: https://t.co/UykHFkRJQl BAIR Blog: https://t.co/DJWhkZMONz 2021-05-04 05:08:24 If you missed @GlenBerseth's talk on SMiRL, catch the poster tomorrow (Tue) at 9 am PT at #ICLR2021. SMiRL is an unsupervised RL method based on *minimizing* surprise -- by aiming to reduce state entropy, SMiRL can produce meaningful behavior in "chaotic" environments. https://t.co/if5aKjBHrS 2021-05-04 02:55:51 I would highly recommend attending Avi's Q&amp 2021-05-04 02:54:20 RT @avisingh599: I'll be answering questions 8.05pm - 8.18pm PT https://t.co/txEQAjrTKX 2021-05-04 00:30:18 RT @avisingh599: I'm presenting our work "Parrot: Data-Driven Behavioral Priors for Reinforcement Learning" at #ICLR2021 right now! If you'… 2021-05-01 22:56:19 @ylecun My brain pattern-matched it to "self-supervised decision transformer" and I got really excited. I don't know what that is, but now I want to write a paper about it. 2021-04-30 19:24:55 That is a great way of putting it -- humans are indeed the rate limiting step. As anyone who has run a robot learning experiment knows, current robot learning is 90% human learning and human labor, and 10% robot effort. We need to change this. Make the robots work harder. https://t.co/nbKVWoi7St 2021-04-29 20:06:15 @antoniloq I think that is a really important goal in the future, because ultimately if we want large datasets, we'll have to pool from many robots. We have a little bit of work in this direction: https://t.co/RGlNrsMWBi But there is a long way to go in this direction. 2021-04-29 17:06:03 I wrote a short, non-technical article summarizing some recent progress in robotic learning, particularly how robots can learn (1) from large offline datasets Robots like the one below: https://t.co/1RwkkydBxC 2021-04-28 16:52:51 New talk covering recent progress on robotic learning from Berkeley and Google -- how can we build data-driven robotic systems with offline RL that learn directly in the real world with minimal human intervention? https://t.co/0J6QfBpfMH 2021-04-28 03:38:03 This generalization of goal-conditioned RL Paper: https://t.co/KzbaNiFR1V Project Website: https://t.co/LVchQkoIGC Poster: https://t.co/MUaM4MeQET YouTube video: https://t.co/lvEnVPy2qX w/ @snasiriany, @Vitchyr, @ashvinair, @AlexKhazatsky, @GlenBerseth 2021-04-28 03:35:34 DisCo RL (distribution conditioned RL) explores this: condition a policy on a distribution, so the policy can capture a wide range of behaviors. E.g., a distribution that has high correlation between positions of two objects induces the policy to put those objects together. 2021-04-28 03:34:13 We can train policies to reach goal states, and then use them to accomplish a range of goals. What if we want policies that can accomplish any task? Distributions over goals can, in theory, represent any task, so we could condition on the parameters of a distribution! a thread&gt 2021-04-28 02:54:32 RT @abhishekunique7: Some cool new updated results for offline pre-training followed by online fine-tuning with AWAC (advantage-weighted ac… 2021-04-27 17:31:47 Pretraining with offline data makes it much easier to get complex real-world robotics tasks, like the dexterous manipulation task above, to work reliably and efficiently. Arxiv: https://t.co/JYyprRInhR Videos: https://t.co/913iRX0ppL w/ @ashvinair, @abhishekunique7, M. Dalal 2021-04-27 17:30:59 AWAC works much like SAC, but with a modified actor update that implicitly enforces a policy constraint, mitigating the OOD action problems that plague offline RL. This works well for *both* offline and online RL, making it great for offline pretraining+finetuning! 2021-04-27 17:30:24 How can we get robots to solve complex tasks with RL? Pretrain with *offline* RL using prior data, and then finetune with *online* RL! In our updated paper on AWAC (advantage-weighted actor-critic), we describe a new set of robot experiments: https://t.co/913iRX0ppL thread -&gt 2021-04-26 16:42:26 w/ @timrudner, @Vitchyr, @rowantmc, @yaringal paper: https://t.co/alF8sGVqU0 2021-04-26 16:42:15 Discount factors and iteratively adjusted shaping over time emerge naturally from our framework, such that the method actually adjusts (automatically) how shaped the reward is while optimizing a valid lower bound. 2021-04-26 16:42:04 This is really nice, because log-probabilities of learned dynamics (like conditional Gaussian models) are much more well-shaped than commonly used goal-conditioned RL rewards, like indicators or epsilon-ball rewards. https://t.co/ToMQDPhiuE 2021-04-26 16:41:47 The main idea: if we treat states and actions as variables in a graphical model, we can derive a variational lower bound on the probability of reaching a state at some point in the future (without knowing when!) and get an RL-like objective where the reward is the *dynamics* https://t.co/rsQhK1fYmt 2021-04-26 16:41:33 Can control as inference give us RL algs that don't need manual rewards? In Outcome-Driven Actor-Critic (ODAC), we show how VI provides principled reward shaping, derive RL as a bound on an inference problem, and show great goal-conditioned results: https://t.co/alF8sGVqU0 &gt 2021-04-23 22:38:42 @TheOneRavenous I just can't make a gif that shows the whole thing (it would be too big) -- the YouTube video on the project website shows the whole thing. 2021-04-23 04:43:07 (this is not actually an exaggeration, we actually had a different commercially available hand, which shall not be named, start smoking after some heavy-duty RL training a few years back) 2021-04-23 04:40:24 Learning algorithms are cool and all, but Vikash's D'Hand design was essential to make this work -- that thing is built like a tank, a lesser robotic hand would be a smoking wreck after this kind of training process https://t.co/Xjea4A9Tb5 2021-04-23 03:59:20 RT @abhishekunique7: We've been working on getting robots to learn in the real world with many hours of autonomous reset free RL! Key idea… 2021-04-23 03:44:18 Paper: https://t.co/sTT9588hC0 Web: https://t.co/iUBV1Cid7K Video: https://t.co/IfXEhwAybN w/ @abhishekunique7, Justin Yu, @tonyzzhao, @Vikashplus, A. Rovinsky, @imkelvinxu, T. Devlin 2021-04-23 03:41:18 This allows a robotic hand+arm to learn a variety of dexterous manipulation skills completely autonomously, practicing by trial and error without any human intervention or oversight. For example, here it is learning a connector task. https://t.co/dNafvYY1re 2021-04-23 03:40:46 The idea in MTRF learning (multi-task reset-free learning) is that, in order to learn complex tasks in the real world, robots should learn many tasks together, where the tasks form a graph such that each task can be "reset" with other tasks. https://t.co/m4QGQYRwiX 2021-04-23 03:40:29 After over a year of development, we're finally releasing our work on real-world dexterous manipulation: MTRF. MTRF learns complex dexterous manipulation skills *directly in the real world* via continuous and fully autonomous trial-and-error learning. Thread below -&gt 2021-04-22 22:11:40 RT @GoogleAI: Today we present a new approach for automated discovery of generalizable #ReinforcementLearning algorithms that evolves a pop… 2021-04-22 05:08:34 I summarized this paper with the following TL It's PRECOG+planning. A cool property that PRECOG has is that it encodes closed-loop behaviors as open-loop sequences of latent codes, and hence can allow for planning of closed loop behaviors easily, which is neat. 2021-04-22 03:17:57 w/ @nick_rhinehart, Jeff He, Charles Packer, Matthew Wright, @mejoeyg, @rowantmc Web: https://t.co/aZTobrV5h5 Arxiv: https://t.co/W7k5yegTIM 2021-04-22 03:16:30 The main idea is to train a latent variable model that represents the intentions of the ego-agent and other agents with latent variables, and instead of planning over actions, plan over the ego-agent's future intentions, while averaging out intentions of other agents. 2021-04-22 03:15:46 Planning in multi-agent settings requires considering uncertain future behaviors of other agents. So we need contingent plans. In our recent work, we tackle this by learning models where open-loop latent space plans lead to closed loop control https://t.co/W7k5yegTIM &gt 2021-04-20 00:25:22 RT @hausman_k: In addition to MT-Opt, we are releasing Actionable Models, which addresses the problem of defining tasks (which becomes quit… 2021-04-19 21:00:23 RT @YevgenChebotar: Excited to present our new work on Actionable Models, an approach for learning functional understanding of the world vi… 2021-04-19 20:14:43 RT @GoogleAI: Presenting two new approaches to robotic #ReinforcementLearning at scale — MT-Opt, an RL system for automated multi-task data… 2021-04-19 20:14:32 RT @hausman_k: Happy to share a project we've been working on for past 2+ years: MT-Opt https://t.co/SyHab7YBkp Multi-task RL at scale on r… 2021-04-19 20:14:24 RT @julianibarz: Our team has reached another milestone towards general robotics: a robot that can learn many tasks at once and be fine-tun… 2021-04-19 20:14:11 RT @hausman_k: Website: https://t.co/OJune1M2Tw Blog: https://t.co/W55VHRpvBx Arxiv: https://t.co/KTVD9cfWzV Done together with a great tea… 2021-04-19 19:34:19 Blog post: https://t.co/rlLTIOgaKf MT-Opt: https://t.co/gweqaiT7oC AMs: https://t.co/aLDnuH9an7 This research was carried out by Google Research, and led by @YevgenChebotar @hausman_k Dmitry Kalashnikov Jake Varley, with contributions from many others! 2021-04-19 19:32:50 In both cases, the resulting multi-task policies can be efficiently finetuned to solve new downstream tasks. In a sense, this provides a kind of general pre-training for robotic policies, learning representations that are effective for a breadth of downstream behaviors. 2021-04-19 19:32:04 Actionable models takes an automated approach, using goal-reaching policies for multi-task training w/o user-specified rewards. By running offline RL, via an algorithm based on conservative Q-learning (CQL), AMs can pretrain goal-reaching policies on huge multi-task datasets. https://t.co/jXf50xbp0P 2021-04-19 19:30:43 Instantiating this requires a few choices -- e.g., how do we define tasks? In MT-Opt, this is done with a set of task reward classifiers, and relabeling that intelligently shares data across different tasks, enabling under-represented tasks to benefit from over-represented ones. https://t.co/W1kZhBWuoI 2021-04-19 19:29:14 The basic principle is simple: gather data for many different tasks across multiple robots, with months of real-world experience, and then use this data to train multi-task policies with offline reinforcement learning. https://t.co/Rd1KSmm01q 2021-04-19 19:28:29 How can we learn diverse multi-task policies via offline RL from huge robotic datasets? In 2 papers (actionable models and MT-Opt), offline RL enables both multi-task pretraining and efficient finetuning to new tasks in real-world settings. https://t.co/rlLTIOgaKf thread -&gt 2021-04-18 01:46:17 Final set of CS182 Deep Learning lectures now added to the course playlist: https://t.co/21GvfFRcHo GANs (Lec 19), adv. examples (20), and meta-learning (21)! More materials on the course website: https://t.co/0PaBOJElo9 2021-04-15 20:11:43 RT @ancadianadragan: assistive typing: map neural activity(ECoG)/gaze to text by learning from the user "pressing" backspace to undo 2021-04-14 22:57:22 RT @_prieuredesion: We're making strides towards truly "in-the-wild" robotic learning systems that can operate with no human intervention.… 2021-04-14 22:57:18 RT @_prieuredesion: Some great demos of exploring unseen cafeterias and fire stations under different seasons and lighting on the project p… 2021-04-14 04:15:08 Check out the full video below Paper: https://t.co/WsEYRIb6P9 Website: https://t.co/Es7YKyv8KX w/ @_prieuredesion @ben_eysenbach @nick_rhinehart https://t.co/tvCAXZvhcs 2021-04-14 04:13:54 The RECON model is trained entirely with offline data. We used data collected over the course of 2020 for several different projects! Offline training across diverse environments allows generalization to lighting, weather, and different settings. https://t.co/uHqK50ZFuM 2021-04-14 04:12:31 As the robot explores a new environment, it builds a "mental map" -- a graph, with edges connected by distances predicted by the learned model. Diverse prior data is used to train the model across many environments, while the graph represents a mental map of the new environment. https://t.co/DXUGKmAf2c 2021-04-14 04:11:30 The idea in RECON is to train a latent goal model, which bottlenecks the goal through a latent variable. By sampling this latent variable from the prior, the robot can explore an entirely new environment (essentially by proposing random goals) much more effectively. https://t.co/ugNYvdmEqo 2021-04-14 04:09:26 Can robots navigate new open-world environments entirely with learned models? RECON does this with latent goal models. "Run 1": search a never-before-seen environment, and build a "mental map." "Run 2": use this mental map to quickly reach goals https://t.co/Es7YKyv8KX &gt 2021-04-09 16:02:53 @zhaomingxie @Mvandepanne For what it's worth -- I agree with you that this article should have done a better job of putting the research in the context of prior work. 2021-04-09 15:57:19 A few folks pointed out to me that there are some factual errors in the MIT TR article. Certainly we are *not* claiming that our paper is the first to show RL for bipedal locomotion! Our prior work section covers lots of prior papers on this: https://t.co/4518eydPzH 2021-04-09 01:48:38 Nice article in MIT Tech Review about our recent paper (@ZhongyuLi4 &amp 2021-04-06 17:07:23 The task reward encodes _what_ we want the character to do, the AMP encodes _how_ we want it to do it. This recipe is general, and I think it will provide a simple and powerful approach to character animation that is much easier to use than state machines, motion graphs, etc. 2021-04-06 17:05:35 This image captures the basic (and simple) idea: take a dataset that describes the desired style of the behavior you want (walks, turns, stylized "zombie walks," whatever), use it to train the motion prior (a discriminator), and then combine it with a task reward. https://t.co/hGpX6OzB2C 2021-04-06 17:03:54 Adversarial motion priors (AMP) provides a simple and scalable data-driven method for physics-based animation: RL is used to train the character to accomplish the task, while data is used to train a "motion prior" that encodes a notion of naturalness. -&gt 2021-04-06 05:23:05 @DhuriHardik Generally I recommend undergraduate level statistics (e.g., a good theory of probability course), or equivalent to Berkeley's CS70, and college-level linear algebra. However, the class goes fast through basic machine learning, so it's better to have taken an intro ML course. 2021-04-06 04:44:43 If you would like to evaluate your off-policy evaluation (OPE) methods, check out our OPE benchmarks. Deep off-policy evaluation (DOPE) -- though you can use it with non-deep methods too if you prefer! https://t.co/31TjOjctdu https://t.co/FCejmT50uw https://t.co/KDXHNw3HgU https://t.co/RDEpJ7z9f5 2021-04-05 04:08:33 Of course, if you want a more in-depth grad-level course on deep RL, you can also check out the full-length CS285 course here: https://t.co/tAGzTDfvUc CS182 covers a broad range of deep learning materials (not just RL), whereas CS285 goes deep (heh) on RL especially. 2021-04-05 04:07:45 I've updated the CS182: Deep Learning with new lectures on RL (policy gradients, actor critic, Q-learning), autoencoders, latent variable models, and VAEs! Website: https://t.co/0PaBOJElo9 Playlist (lectures 15-18): https://t.co/21GvfFRcHo GANs and adv examples coming next! 2021-04-02 23:53:19 @SiddKotwal Yup, we checked this. The github has all the homework materials, and the slides should all be accessible on the website! 2021-04-01 20:39:35 RT @shaneguML: Policy Information Capacity (aka PIC, reward empowerment) is accepted as 1 of 3 contributed talks @iclr_conf Never-ending RL… 2021-03-30 04:44:10 This was a really fun collaboration to get motion imitation + RL for bipedal locomotion with the Cassie robot! Some really exciting experimental results, and a few unintended "emergent" recoveries https://t.co/J4RGtI2OtC 2021-03-29 17:34:18 Paper: https://t.co/uRwMubdlsj Website: https://t.co/e5QrrGRylv @sidgreddy's talk for ICLR: https://t.co/M8jM4oMv1N An awesome collaboration: Jensen Gao &amp @GlenBerseth &amp Nicholas Hardy, @NikhileshNatraj, @KaruneshGanguly (UCSF) 2021-03-29 17:30:42 Our colleagues at UCSF conducted experiments with X2T with a patient who has an ECoG electrode array implanted into their brain, and showed that X2T could allow the patient to type words more effectively using the BCI interface than a non-adaptive baseline learning system! https://t.co/IFvsSX9IiM 2021-03-29 17:29:29 The main idea is that each time the user presses "backspace" to reverse the previous word or character, the system receives negative reinforcement. Over time, it tailors its behavior to the particular user, allowing for human-machine "co-adaptation". Here is a writing example: https://t.co/uQq5nw9gr8 2021-03-29 17:28:37 X2T is a reinforcement learning system that helps users control computers with: gaze, handwriting, and brain-computer interfaces (!). The user performs a task (e.g., typing), while rewarding the interface for correctly parsing their signals. -&gt https://t.co/e5QrrGRylv https://t.co/Ez4zqdC0bC 2021-03-25 23:14:53 @jmac_ai @abhishekunique7 All the same tricks for Q apply as far as I know, though we didn't study in great detail precisely which of these tricks are most important in this case, though maybe @ashvinair has other thoughts on this? 2021-03-25 03:12:31 RCE enables RL to learn behaviors from data -- RCE does not require a reward function, and does not attempt to learn one, but directly learns policies that match a distribution of user-provided success states. It generalizes C-Learning to arbitrary tasks https://t.co/ieEmqI8aFq https://t.co/HPjVaYexks 2021-03-20 21:18:13 @SiddKotwal Can you let me know what you were trying to access that requires a UCB ID? The homework github repos appear to be accessible to everyone. 2021-03-15 03:24:18 @ChrisDeCstVerde Yes, CS285 homeworks are linked here: https://t.co/tAGzTDfvUc For CS182, we currently have HW1 on the website: https://t.co/0PaBOJElo9 We'll make the rest of the CS182 homeworks available on the above website shortly. 2021-03-13 23:23:17 Note that this is separate from my deep RL course -- CS182 covers deep learning broadly (only a few lectures about RL, contrary to the typo I had in the description...). If you want to go deep (heh) on RL in particular, check out my other course: https://t.co/SzCxe71WVB 2021-03-13 23:20:05 Big thanks to the following courses and faculty that I borrowed material and curriculum ideas from: Andrej Karpathy et al. (CS231n) Chris Manning (CS224n) John Canny (prior offerings of CS182) I did my best to credit whenever possible, but I probably missed many, sorry! 2021-03-13 23:18:13 Topics include: - basic machine learning - convolutional neural networks - recurrent neural networks - transformers - NLP - reinforcement learning - generative models &amp (so far we've covered everything up to RL, which will be posted in a few weeks) 2021-03-13 23:17:22 I'm releasing all the lectures (so far) for my deep learning class, CS182! This is an introductory deep learning course (advanced undergraduate + graduate) covering a broad range of deep learning topics. Website: https://t.co/0PaBOJElo9 Playlist: https://t.co/21GvfFRcHo -&gt 2021-03-13 04:14:16 @Vikashplus @chelseabfinn Yeah, the model is actually a badly butchered version of the PR2 model that I believe you converted from the WG URDF. 2021-03-12 21:01:14 As @chelseabfinn just pointed out to me, there is even proof (complete with awkward comment from me) https://t.co/XnikaLRIzS 2021-03-12 18:50:48 Also, here is a cool animation of a robust (stochastic) policy pushing an object around a barrier Ben did all the work here, but I can claim a small contribution: in 2014 I made the robot model, and in 2015, on suggestion of Marvin Zhang and @chelseabfinn, I added the eyes... https://t.co/A3v3UCTKMh 2021-03-12 18:41:31 But precisely *how* MaxEnt RL is robust (ito what set of perturbations) has been a mystery. This work expands our earlier work on how MaxEnt RL is robust to *reward* perturbations: https://t.co/xOdhcvO19J But we now show dynamics robustness, which is arguably much more useful. 2021-03-12 18:40:03 Of course, relationships between maximum entropy models and robustness in general are widely known, and the robustness perspective is in fact the origin of maxent modeling in the literature. Hence people suspected for a long time that MaxEnt *RL* must be robust in some way. 2021-03-12 18:39:04 The particular stochasticity induced by MaxEnt RL results in an objective that lower-bounds a min-max robust control objective, meaning that policies learned by MaxEnt RL should be provably robust to certain perturbations -- this is not the case for regular RL. 2021-03-12 18:37:47 We often hear that MaxEnt RL is "robust" -- but what does that mean? In his new paper &amp -&gt Paper: https://t.co/LJs50mwdLb Post: https://t.co/j5X2MW6GcK 2021-03-11 05:08:12 RT @shiorisagawa: We’ve released v1.1 of WILDS, our benchmark of in-the-wild distribution shifts! This adds the Py150 dataset for code comp… 2021-03-09 19:15:08 What is offline reinforcement learning? I made a talk giving a *non-technical* overview that explains offline RL. Since these days we pre-record our talks, I figured I would also share it with all of you! If in a hurry, watch at 2x speed... https://t.co/jJxanVSV5X 2021-03-05 23:50:41 New blog post from @natashajaques &amp https://t.co/7vTTsA9RKS https://t.co/pjeZgNJjUz 2021-03-01 20:47:04 Great article by @TiernanRayTech on robot learning, including our work at Berkeley &amp Hopefully will encourage even more researchers to tackle the difficult problems in robotic RL! 2021-02-17 06:12:19 RT @chelseabfinn: A new algo for RL from offline data: COMBO pushes down the est. value of states &amp 2021-02-17 06:06:37 Also, I couldn't be more proud of such a wonderful group of collaborators for this project: @TianheYu, @aviral_kumar2, @rmrafailov, @aravindr93, @chelseabfinn 2021-02-17 05:33:25 The reason I am really excited about this paper: the empirical results are very promising, for both state and image-based tasks, and the theory provides us with some insight about why this works. This gives us good confidence that the lessons will generalize to other tasks too. 2021-02-17 05:20:17 A principled offline model-based RL method: COMBO combines conservative Q-learning (CQL) with model-based learning, providing state-of-the-art offline RL results and formal guarantees! https://t.co/qi3u9erxOV w/ @TianheYu, Aviral Kumar, R. Rafailov, @aravindr93, @chelseabfinn https://t.co/IFGbStG2Oc 2021-02-08 21:52:05 RT @julianibarz: Our latest paper summarizing what we learned for the last 5+ years when applying RL to robotics. https://t.co/SbXppEXH9j 2021-02-08 19:12:17 This project was led first and foremost by @julianibarz, with a wonderful group of colleagues who contributed portions of the paper: Jie Tan, @chelseabfinn, Mrinal Kalakrishnan, Peter Pastor. 2021-02-08 19:11:25 It is important in robotic RL to think about not just the math in the algorithms, but the practicalities of getting learning systems to work in the real world: ops (can the robot keep training for a long time), safety, resets, scalability, etc. 2021-02-08 19:10:49 It's also a little bit out of date at this point (it's a journal paper, which took nearly a year to clear review, despite having very few revisions... but that's life I suppose). But we hope it will be pretty valuable to the community. 2021-02-08 19:09:37 This is somewhat different from the usual survey/technical paper: we are not so much trying to provide the technical foundations of robotic deep RL, but rather describe the practical lessons -- the stuff one doesn't usually put in papers. 2021-02-08 19:09:01 What did we learn from 5 years of robotic deep RL? My colleagues at Google and I tried to distill our experience into a review-style journal paper, covering some of the practical aspects of real-world robotic deep RL: https://t.co/fYGQfFYlKu -&gt 2021-02-03 20:20:45 RT @doomie: Blog post (with @babaeizadeh) describing our work and the library that can create some of the magic! https://t.co/ClqOtrUUuz 2021-02-03 20:20:38 RT @babaeizadeh: Blog post on our latest experiments on visual model based reinforcement learning https://t.co/XlS1aME66z One of the most… 2021-01-18 19:00:30 @pyoudeyer @FlowersINRIA It's inspiring how simple, cheap robots can be used to do cutting-edge research like this. For intrinsic motivation, lifelong learning, and other "persistent exploration" things, it's a great formula. Also safer and more accessible than bulky, expensive industrial robots :) 2021-01-10 22:25:36 Covers these papers: AWAC: https://t.co/JYyprRInhR MOPO: https://t.co/53VtOZKbcx CQL: https://t.co/TYL7RTNbO7 COG: https://t.co/i4rVXZoQbF 2021-01-10 22:24:25 This was a fun talk! An extended version of my offline RL + common sense talk, also covering model-based offline RL. https://t.co/6I2kADwdTD 2021-01-10 01:24:37 On a personal note, this was a really enjoyable collaboration with an amazing team: Stephen Tian, @SurajNair_1, Frederik Ebert, Sudeep Dasari, Ben Eysenbach, @chelseabfinn (Berkeley, Stanford, CMU) Couldn't have wished for a better group of collaborators for this :) 2021-01-10 01:22:35 This outperforms a range of prior methods in simulation, and can even control a real-world robot to open/close a drawer! (robot experiments were done by Suraj Nair, in @chelseabfinn's lab at Stanford) https://t.co/yV3ggQ5ai4 2021-01-10 01:21:28 The sampling based (CEM) planner then rolls out sampled action sequences, and selects the one that ends up with the shortest (predicted) dynamical distance to the goal. Learning the distance function turns out to be just an offline Q-learning problem. https://t.co/LP9x22VEzU 2021-01-10 01:20:42 The idea: we learn a predictive model (predict what will happen next given image+action), but this is only good for short(ish) horizon planning. For longer horizons, we also learn a "dynamical distance" that predicts how many steps to the goal, which we use as a planning cost https://t.co/PpOZXAxRSU 2021-01-10 01:19:52 Offline model-based RL for goal reaching: learn a distance "Q-like" function from offline data, and a video prediction model, then use them to accomplish visually indicated goals. w/ Stephen Tian et al. https://t.co/pmXL8fGHXv https://t.co/x9XXI7PN06 &gt 2021-01-07 03:05:55 My talk from the Knowledge Based RL Workshop at IJCAI can be watched here: https://t.co/xB3csuo9zw The talk covers COG and CQL, with a perspective on how offline RL can implicitly acquire knowledge and "common sense" from prior data! 2021-01-07 02:28:28 At IJCAI? I'm giving a talk in the Knowledge Based RL Workshop at 6:30 pm PT (in 5 min!), on how offline RL can distill knowledge &amp 2021-01-06 20:01:18 How can we train generative models to predict possible future outcomes, without regard for time steps, and then use them for control? @michaeljanner's BAIR blog post on gamma-models explores this question, with some neat visualizations! https://t.co/KnzYVERXyq 2020-12-18 02:27:41 I think this method is promising for a few reasons: - offline training using large prior datasets - fully autonomous - flexible, new goals specified just using images In experiments, data was collected ~6 mo prior (for another project), so it's even robust to changing seasons :) 2020-12-18 02:26:14 Or you could command the robot to patrol an area by giving it visually indicated waypoints, simply by snapping photographs of each waypoint. 5/n https://t.co/s4AmMafOSK 2020-12-18 02:25:30 This system can then be used in a few interesting ways: we can define "contactless delivery" targets just by having someone take a photo of their front door, and the robot then navigates to their front door to deliver their package. 4/n https://t.co/r8gIKBk1qR 2020-12-18 02:24:44 Once we have a distance function, policy, and graph, we search the graph to find a path for new visually indicated goals (images), and then execute the policy for the nearest node. A few careful design decisions (in the paper) make this work much better than prior work. 3/n https://t.co/45BR3xCsCi 2020-12-18 02:23:25 The idea: use RL + graph search to learn to reach visually indicated goals, using offline data. Starting with data in an environment (which in our case was previously collected for another project, BADGR), train a distance function and policy for visually indicated goals. 2/n https://t.co/2HtNr6N894 2020-12-18 02:21:31 RL enables robots to navigate real-world environments, with diverse visually indicated goals: https://t.co/r6m5yJYrQW w/ @_prieuredesion, B. Eysenbach, G. Kahn, @nick_rhinehart paper: https://t.co/MRKmGStx6Y video: https://t.co/RZVVD2pku7 Thread below -&gt 2020-12-17 20:15:53 Aviral Kumar and I have posted our NeurIPS offline reinforcement learning tutorial on YouTube for your enjoyment :) Slides, colab exercise, etc.: https://t.co/S639WkAroh Part 1: https://t.co/OozPaXLVhF Part 2: https://t.co/MPLhyipS1K 2020-12-17 20:14:03 All of the main lectures for UC Berkeley's fall 2020 deep RL course are now posted: https://t.co/SzCxe71WVB Newly posted lectures: 21 (transfer learning), 22 (meta-learning), 23 (open problems)! 2020-12-16 02:34:45 Distribution shift is a fact of life for models in the real world: the test distribution will never really match the training distribution. But evaluating methods that account for distribution shift is difficult. WILDS aims to address that, across a range of realistic domains. https://t.co/WNP1nRUXvb 2020-12-15 18:48:50 @rowantmc @ToyotaResearch @adnothing Congrats Rowan on your new job, we'll miss you! 2020-12-15 05:21:36 This online setting is I think an especially good fit for meta-learning: the final model after seeing many tasks can generalize in zero-shot, but long before it gets there, meta-learning can accelerate how few tasks/images are needed to get to that point! 2020-12-15 05:20:50 One neat thing that this method can do is open-domain classification: a new class is added with each task, and the network adapts by gradient descent, using a Siamese architecture. Learning is online, the classifier for each class must improve after every single new datapoint. https://t.co/ZDWO1YR5ps 2020-12-15 05:19:17 Online meta-learning methods learn to learn continually, becoming better at learning new tasks with each task. In new work, we demonstrate incremental online meta-learning, which learns one datapoint and task at a time @TianheYu, X. Geng, @chelseabfinn https://t.co/LWbH2hUHnB https://t.co/PtwlTNR5Pp 2020-12-12 18:53:00 @unsorsodicorda So far it does seem to work quite well, though the tasks are a bit different (eg we didn’t test it on images). We are still finalizing the paper, so this is a “fresh off the press” kind of thing (not on arxiv yet, just at the workshop) 2020-12-12 00:39:06 At WS on challenges of real-world RL, B. Eysenbach will present DARC. Domain adaptation for RL: how to train RL agents in one domain, but have them pretend they are in another Pres (1520 PT Sat 12/12) https://t.co/8eczV4hkM8 Paper https://t.co/FqoaM5BxmS https://t.co/Sa9jbozns4 https://t.co/0UiUZlw073 2020-12-11 21:18:14 At ML4Molecules (https://t.co/yyLUzGsRW6), Justin Fu will present Offline Model-Based Optimization via Normalized Maximum Likelihood (NEMO), for optimizing designs from data w/ NML! 8:30am ML4Molecules poster session Paper: https://t.co/N9jWfIZhOS Poster: https://t.co/ZaYycpuhAF https://t.co/gwo1iteH0Q 2020-12-11 20:16:37 RT @agarwl_: Today, we will be presenting our work on how deep neural networks interact with Q-learning (w/t Aviral, @its_dibya, @svlevine)… 2020-12-11 18:43:25 At DRL WS and robot learning WS, @_oleh, @chuning_zhu will present collocation-based planning for image-based model-based RL! By relaxing dynamics, robot images object "flying" to goal before figuring out how to move it Video https://t.co/4ePoBMUvkT paper https://t.co/g8hp7bLas2 https://t.co/ksEfhdJ6V8 2020-12-11 18:36:38 Also at deep RL WS, Abhishek Gupta &amp Room D, C7, Deep RL Workshop, 12:30-1:30 and 6-7 PST Paper: https://t.co/gA6HhPNVqy Slideslive: https://t.co/Q5hE42riUa https://t.co/2lu935FeYT 2020-12-11 18:35:38 At deep RL WS, Abhishek Gupta &amp Room B, B5, 1230-1330 &amp Paper https://t.co/fn8sEiLOsT Pres https://t.co/FO7UwlZ53k Blog https://t.co/8wZp0pEiOx https://t.co/QUlaG5lTZf 2020-12-11 05:32:47 Enjoy all the NeurIPS workshops!! 2020-12-11 05:32:23 At meta-learning workshop, Marvin Zhang will present Adaptive Risk Minimization, how models can learn to adapt to distributional shifts at test time via meta-learning. Paper: https://t.co/u1FZBiUJZ1 Pres: https://t.co/T8bRr8N0UC https://t.co/KvfQubyjEz 2020-12-11 05:31:35 At robot learning WS (8:45am PT poster) and real-world RL WS (12/12 11:20am poster), @avisingh599 will present PARROT: pre-training models that explore for diverse robotic skills. Arxiv: https://t.co/COMGTmCInG Video: https://t.co/2MU2V8VNGa Website: https://t.co/1o1T6rsiHb https://t.co/sttRDBgGp8 2020-12-11 05:31:11 At deep RL WS, and as long oral presentation at offline RL WS, @avisingh599 will present COG: how offline RL can chain skills and acquire a kind of “common sense” Vid: https://t.co/v0yt1j8Lh2 Web: https://t.co/6A9INvrB5G Blog: https://t.co/cAhxunkbAN Offline RL talk 12/12 9:50am https://t.co/EY6AitWuF9 2020-12-11 05:30:42 At deepRL WS, @mmmbchang will present “Modularity in Reinforcement Learning: An Algorithmic Causality Perspective on Credit Assignment” how causal models help us understand transfer in RL! Poster: https://t.co/85qJafdpAg Paper: https://t.co/RIuQchAnjI Vid: https://t.co/ELHZEtn0Eg https://t.co/xspMNqNPyc 2020-12-11 05:29:53 Also at deep RL WS, @snasiriany and co-authors will present “DisCo RL”: RL conditioned on distributions, which provides much more expressivity than conditioning on goals. Paper: https://t.co/SYJqtfHfoT Presentation: https://t.co/c2zGwg8q8u Poster: https://t.co/MUaM4MeQET 2020-12-11 05:27:58 Also at deep RL WS, Aviral Kumar will present “Implicit Under-Parameterization” – our work on how TD learning can result in excessive aliasing due to rank collapse. Paper: https://t.co/haeE1YX4Ue Video: https://t.co/xoEEq2t0Gh https://t.co/Z7nddVBV3z 2020-12-11 05:27:35 Also at deep RL WS posters, @timrudner @vitchyr will present “Outcome-Driven Reinforcement Learning,” describing how goal-conditioned RL can be derived in a principled way via variational inference. Paper: https://t.co/aaG6wm2Kvp Talk: https://t.co/cWS0dH1fIN https://t.co/n0iRAmtUxG 2020-12-11 05:26:24 Ashvin Nair will present AWAC, offline RL with online finetuning, also at the deep RL WS poster session. pres: https://t.co/Ns9kDNsG90 paper: https://t.co/JYyprRInhR blog: https://t.co/4MAkic9mT4 https://t.co/NFhsjmtOSz 2020-12-11 05:26:00 Also at deep RL WS posters, Jensen Gao&amp Paper: https://t.co/ygkCJH8pmJ Talk: https://t.co/60T8YKh3TO https://t.co/nyHZ0dhpee 2020-12-11 05:24:43 Ben will also present C-Learning: a new algorithm for goal-conditioned learning that combines RL with principled training of predictive models. Deep RL poster session, 12:30 pm PT. Paper: https://t.co/Ms0e9wvxaA Website: https://t.co/GUhYjmAIXC Talk: https://t.co/5G30WgNZcy https://t.co/lOSoxSf97V 2020-12-11 05:24:25 At the deep RL workshop, Ben Eysenbach will talk about how MaxEnt RL is provably robust to certain types of perturbations. Contributed talk at 14:00pm PT 12/11. Paper: https://t.co/ZwaGqzwdtb Talk: https://t.co/BvjCFW257j 2020-12-11 05:24:13 At robot learning workshop, @katie_kang_ will present the best-paper-winning (congrats!!) “Multi-Robot Deep Reinforcement Learning via Hierarchically Integrated Models”: how to share modules between multiple real robots 2020-12-11 05:23:28 My favorite part of @NeurIPSConf is the workshops, a chance to see new ideas and late-breaking work. Our lab will present a number of papers &amp thread below -&gt meanwhile here is a teaser image :) https://t.co/OCiU537VPH 2020-12-10 21:14:53 In-depth analysis of image-based model-based RL by @babaeizadeh and colleagues is finally out! Lots of analysis studying why model-based RL with images works, and which design decisions are more or less important. Some of the conclusions may surprise you! (or they might not...) https://t.co/dhHbLOM1Ec 2020-12-10 21:13:20 @kargarisaac Yes, or get labels somewhere (e.g., human labelers). Nice thing with offline data (whether MBO or offline RL) is that the labels only need to be obtained once, just like with any supervised learning application. 2020-12-10 21:12:04 @unsorsodicorda We do have experiments with real-world datasets (images, proteins, etc.). Curious what kind of applications you would have in mind -- we are actively trying to find suitable domains in which to evaluate these methods. 2020-12-10 19:36:20 In contrast, many alternative methods rely on active sampling of the data -- when we are talking about real world data (e.g., biology experiments, aircraft designs, etc.), each datapoint can be expensive and time-consuming, while offline MBO can reuse the same data. 2020-12-10 19:35:36 Why does this problem matter? In many cases we *already* have offline data (e.g., previously synthesized drugs and their efficacies, previously tested aircraft wings and their performance, prior microchips and their speed), so offline MBO uses this data to produce new designs. 2020-12-10 19:31:43 MINs address this issue with a simple approach: instead of learning f(x) = y, learn f^{-1}(y) = x -- here, the input y is very low dimensional (1D!), making it much easier to handle OOD inputs. This ends up working very well in practice. 2020-12-10 19:31:06 This can be mitigated by active sampling (i.e., collecting more data), but this is often not possible in practice (e.g., requires running costly experiments). Or by using Bayesian models like GPs, but these are difficult to scale to high dimensions. 2020-12-10 19:30:30 Classically, model-based optimization methods would learn some proxy function (acquisition function) fhat(x) = y, and then solve x* = argmax_x fhat(x), but this can result in OOD inputs to fhat(x) when x is very high dimensional. 2020-12-10 19:29:52 The problem setting: given samples (x,y) where x represents some input (e.g., protein sequence, image of a face, controller parameters) and y is some metric (e.g., how well x does at some task), find a new x* with the best y *without access to the true function*. 2020-12-10 19:29:08 Tonight 12/10 9pm PT, Aviral Kumar will present Model Inversion Networks (MINs) at @NeurIPSConf. Offline model-based optimization (MBO) that uses data to optimize images, controllers and even protein sequences! paper: https://t.co/ePP7yjAgj0 pres: https://t.co/lsJ1qreCmj more-&gt 2020-12-10 02:49:58 Tmrw (12/10), @imkelvinxu will present "Continual Learning of Control Primitives: Skill Discovery via Reset-Games": lifelong learning without resets, while acquiring a repertoire of skills Check it out at the 9am PT poster session at @NeurIPSConf! Paper: https://t.co/EtNpMR2dy8 https://t.co/xK8uxQ3Tq8 2020-12-09 22:27:59 Come check out Tianhe Yu's PCGrad presentation at @NeurIPSConf, 9pm PT! Links below PCGrad mitigates conflicting gradients in multi-task learning, with a simple project trick that cancels off conflicting portions of gradients, greatly improving performance in multi-task settings https://t.co/UbB6SmJaao 2020-12-09 02:42:24 CQL will be presented in the 9 pm PT poster session at @NeurIPSConf on 12/8 (Tuesday, today). Paper: https://t.co/TYL7RTNbO7 Website: https://t.co/u0eVrncvAF Poster session: https://t.co/mH7c5qi7vC 2020-12-09 02:39:49 Today at 9pm PT (in 2.5 hrs), Aviral Kumar will present Conservative Q-Learning (CQL). CQL learns Q-functions that lower bound the true Q-function, providing some interesting theoretical guarantees, as well as excellent empirical performance. Come by Aviral's poster! more below&gt 2020-12-08 18:09:51 Paper: https://t.co/g0VQv7GlBS Presentation (6 pm PT today 12/8): https://t.co/U11qNs6C2H w/ Ben Eysenbach, Young Geng, @rsalakhu 2020-12-08 18:09:15 The idea: for all collected, infer the posterior over tasks for which that behavior is near-optimal. This is the same as the inverse RL problem. Then use this posterior to relabel this experience. This can work for any reward class: goals, linear rewards, non-linear rewards, etc. 2020-12-08 18:08:25 Today (Tue 12/8) check out Ben Eysenbach's long talk on HIPI: Rewriting History with Inverse RL at @NeurIPSConf, at 6 pm PT: https://t.co/U11qNs6C2H Find out how inverse RL allows us to formalize relabeling of rewards, improving learning of contextual policies. more-&gt 2020-12-08 17:28:04 RT @chelseabfinn: Interested in robustness in reinforcement learning? @SaurabhKRL is presenting this paper at @NeurIPSConf right now (9-11… 2020-12-08 05:52:54 And one more for tomorrow: Alex Lee and Anusha Nagabandi will present SLAC: Stochastic Latent Actor-Critic. Come learn how representation learning can drastically speed up RL! 9 am PT Tue 12/8 Pres: https://t.co/lwHkiXnLrC Web&amp 2020-12-08 05:00:12 Tmrw @NeurIPSConf, @michaeljanner will present gamma-models at 9am PT (12/8): models that predict densities over potential future states over a long horizon, rather than one step at a time. Pres: https://t.co/CwOWRKiCFr Paper: https://t.co/j4FKAQmrWK Web: https://t.co/QYga0n3Vy4 https://t.co/pSwDoa2F6n 2020-12-08 04:54:37 HEDGE trains "tree-structured" models that can generate intermediate subgoals between pairs of states. Paper: https://t.co/GlSlAEuvR0 Website: https://t.co/Rre1xTzu55 Video presentation: https://t.co/XGAFhjJVfw 2020-12-08 04:53:42 Tmrw (Tue 9/8 at 9 am PT) check out HEDGE at @NeurIPSConf: hierarchical planning with learned tree-structured models enables planning complex behaviors one subgoal at a time. w/ @KarlPertsch, @_oleh, @febert8888, @chelseabfinn, @dineshjayaraman https://t.co/nODdDelyKP more-&gt 2020-12-08 02:06:05 This is going to be presented tonight (in 3 hrs) at @NeurIPSConf, 9 pm PT poster session! Come find out about offline model-based RL. https://t.co/OGf7J27O1b 2020-12-07 22:14:03 RT @berkeley_ai: Check out Aviral Kumar and Avi Singh's @avisingh599 new blog post, "Offline RL: How Conservative Algorithms Can Enable New… 2020-12-07 19:06:05 The work in this blog post will also be presented at @NeurIPSConf! CQL will be presented Tue 9 pm PT in the poster session COG will be presented at the offline RL workshop in the 12:40 pm PT contributed talks session 2020-12-07 19:03:03 Want to learn even more offline RL? Aviral &amp Offline Reinforcement Learning: How Conservative Algorithms Can Enable New Applications Overview of CQL &amp 2020-12-07 16:20:14 Our @NeurIPSConf tutorial on offline RL has started! Aviral and I are answering questions live via chat during the presentation, with a Q&amp Website with colab: https://t.co/S639WkAroh 2020-12-07 04:40:04 @SketcherRami @NeurIPSConf I posted the wrong time, but I believe I deleted the previous one, so this should be the correct done 2020-12-07 02:20:23 Paper: https://t.co/2LxmUc6b6J Presentation: https://t.co/6eVsfbVCtm Poster: https://t.co/jMoghV3WdU Blog Post: https://t.co/aU6EGeG3zs Poster session: 9-11pm PT Monday Spotlight talk + Q&amp w/ Aviral Kumar, @abhishekunique7 2020-12-07 02:20:14 Can we stabilize deep RL by only backing up target values that are more likely correct? DisCor aims to achieve this, providing for "corrective feedback" during learning. See Aviral's @NeurIPSConf talk tmrw (Mon), 7:20pm PT: https://t.co/6eVsfbVCtm (poster at 9pm) more below -&gt 2020-12-06 22:52:24 This is joint work with: @MichaelD1729, @natashajaques, @EugeneVinitsky, @alexandrebayen, Stuart Russell, Andrew Critch paper: https://t.co/JCOUElGUTv code: https://t.co/jUWFGYZDsc 2020-12-06 22:51:33 Adversarial environment design: PAIRED learns to generate complex environments via minimax regret. Check out our presentation at @NeurIPSConf: Oral: 6:30 pm PT Mon 12/7: https://t.co/aLK2dbjzP2 Poster: 9 pm PT Mon: https://t.co/aLK2dbjzP2 more below: https://t.co/yDfoj5gI3k https://t.co/3IUMfQ1556 2020-12-05 23:06:10 There are also some more materials on the tutorial website: https://t.co/S639WkAroh Including a colab demo/exercise authored by Aviral for you to play with offline RL in gridworld domains, including an illustrative CQL implementation: https://t.co/QssLGJjVBc 2020-12-05 22:32:49 If you prefer to learn by reading rather than watching, you can also read our tutorial and survey paper on offline RL here: https://t.co/kv4mtVMScG If you want to try out some offline RL problems, try D4RL, the most popular offline RL benchmark suite: https://t.co/dPMs20puRW 2020-12-05 22:31:47 Want to learn about offline RL? Aviral Kumar and I are giving a tutorial on offline RL at @NeurIPSConf. Materials now online! https://t.co/ka741RCymv Main talk: 8 - 10:30 am PT 12/7 Mon Q&amp Recording here: https://t.co/24HFYHiAE2 more -&gt 2020-12-03 04:43:08 @JeffDean @doomie @eigenhector Mine was about animating octopus hand gestures. 2020-11-21 00:34:21 What do pigeons have to do with robots? Both can learn with offline RL and combine skills from prior experience! Find out more about pigeons &amp https://t.co/v0yt1j8Lh2 2020-11-20 17:24:52 I posted the paper link a few days ago for this, here it is again: https://t.co/piwcU5wMs4 High level idea: train on test point with each label, and see which model gives the highest likelihood. Making this tractable with amortization (ACNML) makes for a useful algorithm. 2020-11-20 17:23:57 Training on test points is a no-no. But training on test points with fake labels turns out to be a good idea. Aurick's blog post on amortized conditional normalized likelihood (ACNML) shows how NML provides a principled view of "training on test points": https://t.co/iasW5xhaQs https://t.co/NkU7Ntl9fi 2020-11-20 15:43:21 I'm excited about this work because the exploration enabled by Parrot really looks much more intelligent than what we get with random movements, and evaluating on diverse tasks tests generalization. w/ @avisingh599, Huihan Liu, Gaoyue Zhou, Albert Yu, @nick_rhinehart 2020-11-20 15:42:28 We train a multimodal density model (based on normalizing flows) on past data, obtained from a variety of tasks, which allows a robot to learn basic principles like reaching for objects and picking things up. When faced with a new scene, robots will reach for random objects. https://t.co/06agbo7Onv 2020-11-20 15:40:58 RL agents explore randomly. Humans explore by trying potential good behaviors, because we have a prior on what might be useful. Can robots get such behavioral priors? That's the idea in Parrot. arxiv https://t.co/COMGTmCInG web https://t.co/1o1T6rsiHb vid https://t.co/2MU2V8VNGa https://t.co/cZSxqTHafl 2020-11-19 20:02:19 What I find exciting about this work is that it provides principles for a variety of "hacks" that are used for goal-conditioned policies, explaining "Q-value-like" quantities as densities, explaining relabeling, and what relabeling ratio to use. Website: https://t.co/GUhYjmAIXC 2020-11-19 20:01:28 C-Learning learns goal-conditioned policies using classifiers, without any hand-designed rewards. The theoretical framework in C-Learning helps to explain connections between prediction and Q-functions, relabeling, and others. https://t.co/tGEGEF52Ga w/ Eysenbach, @rsalakhu https://t.co/z3co465aBc 2020-11-18 19:45:47 @CsabaSzepesvari @neu_rips @jacobmbuckman Though I guess these days all of us are sitting at home by ourselves, so maybe the "maverick scientist" mythos feels more true now than it did before 2020-11-18 19:43:20 Today at 11:50 am (in 10 min!) @ryancjulian will present (at @corl_conf) "Never Stop Learning" -- finetuning for robotic RL. Come learn how finetuning enables large-scale robotic RL to adapt. talk https://t.co/o5uU7MMzTS CoRL https://t.co/5bp9Nptcww paper https://t.co/8DlOHcDrGA https://t.co/BRSQcE5H1B 2020-11-18 01:02:24 And here is a bonus animation visualizing the exploration process :) Presentation Link: https://t.co/X5LCIwynpi Paper: https://t.co/tJkg3RM9sE https://t.co/bWvKEtKUJW 2020-11-18 01:02:05 On Wed 11:10 - 11:40 AM at @corl_conf, we will present MELD, meta-RL from images for real-world robotic manipulation! Come find out how to get robots to figure out where to insert ethernet cables from images. w/ Tony Zhao, K. Rakelly, A. Nagabandi, @chelseabfinn https://t.co/UHxgNlHFqJ 2020-11-17 23:03:54 @jacobmbuckman @neu_rips I do think there are several huge problems, I just don't think I'm the only one who is aware of them. Science is a social endeavor, important discoveries from maverick scientists who work in a dark basement are a myth. 2001-01-01 01:01:01