Sergey discusses several important questions (1) describing the RL task to an agent, (2) using prior information to generalize (3) bootstrapping using _behavioral priors_ and (4) fully automating tasks. Although this talk is limited to robotic RL (natural, since that's his area of expertise), answering these questions is clearly crucial to progress in this field. One question not touched upon in this talk concerns *teams of robots* . Going back to the Robinson Crusoe analogy, a _team_ of robots, *specializing* in different tasks would stand a much better chance of surviving. One shouldn't brush this question aside saying it's hard enough to train one robot let alone an entire team. Recent results in MARL are showing emergent behaviors in robotic teams that are truly unexpected and quite surprising. I'm just saying the community should probably also investigate what these questions (and their solutions) look like in a MARL setting soon.
Inspiring Presentation! Real-World Reinforcement Learning might allows to develop policies generalized enough to be the building blocks of artificial common sense.
Fantastic talk. Many thanks, Sergey.
Great presentation I really enjoyed it
Sergey discusses several important questions (1) describing the RL task to an agent, (2) using prior information to generalize (3) bootstrapping using _behavioral priors_ and (4) fully automating tasks. Although this talk is limited to robotic RL (natural, since that's his area of expertise), answering these questions is clearly crucial to progress in this field.
One question not touched upon in this talk concerns *teams of robots* . Going back to the Robinson Crusoe analogy, a _team_ of robots, *specializing* in different tasks would stand a much better chance of surviving. One shouldn't brush this question aside saying it's hard enough to train one robot let alone an entire team. Recent results in MARL are showing emergent behaviors in robotic teams that are truly unexpected and quite surprising.
I'm just saying the community should probably also investigate what these questions (and their solutions) look like in a MARL setting soon.
Excellent and very interesting talk. Thank you.
nice to see that you ask the right questions or at least better than a lot of RL researcher good job
Great talk thank you
Inspiring Presentation! Real-World Reinforcement Learning might allows to develop policies generalized enough to be the building blocks of artificial common sense.