ImageNet moment for robot locomotion. Congratulations et al!
"first, we train a policy using RL with a cheap-to-compute variant of depth image and then in phase 2 distill it into the final policy that uses depth using supervised learning."
vision-locomotion.github.io
Conversation
Replying to
Such behavior engineering + distillation paradigm is the only approach that scales, to achieve generalization in deep RL. We have a new upcoming (proof-of-concept) paper soon!
Quote Tweet
Quite aligned with our recent paper decomposing deep RL to "behavior generation" and "behavior distillation" arxiv.org/abs/2110.04686 ...
Show this thread
1
5
23
Quote Tweet
Why? See language-play.github.io (and our arxiv.org/abs/1906.07343). Once you have rich behaviors, you can distill to a policy with any *observation type* (e.g. language or images, esp in sim) or *policy architecture*. This supervised part with no RL can be explored by 100x people
Show this thread
1
3
Replying to
These four people are to Boston Dynamics, what Alex, Ilya, Geoff were to image recognition. 3 yrs of work by a tiny team, but an approach completely scalable with compute. GPT moment for robot locomotion will come.
3
3
45
Funny thing: I remember the debate between Geoff and Jitendra in 2012, where I recall Jitendra appeared still a bit skeptical of the implication of ImageNet to the rest of CV. Now Jitendra is doing similar thing to robotics community :)
1
1
6
Show replies
Replying to
Jitendra's response: "Thanks! Yes, indeed I think of this as a big win for RL, and training in simulation as a way forward for robotics. ..."
1
2
2
"This transfers the problem to the "sim-to-real" stage, for which we presented a solution "Rapid Motor Adaptation" at RSS 2021 (you can think of it as the classic problem of "adaptive control" done in the framework of deep learning)"
1
1
2
Show replies
Likely the biggest RL successes in the last few years if not imagenet, so I'll give you that.
2