Opening Challenges & Open Problems in Deep Reinforcement Learning
Challenges & Open Problems in Deep Reinforcement Learning
- Length
- 38 min
- Level
- Graduate
- Recorded
- UC Berkeley2023
Chapter 1 of 4: Reinforcement learning as an engineering tool
Deep reinforcement learning faces open problems that Sergey Levine frames through three readings of the field in UC Berkeley's CS 285. The first treats reinforcement learning as an engineering tool that inverts a physical simulator into a control law. Moravec's paradox, set against Deep Blue and AlphaGo, then motivates agents that learn in the physical world. Reset-free multi-task dexterous manipulation turns failures into practice, and Parrot's behavioral priors bootstrap exploration. The third reading casts reinforcement learning as a framework for decisions from low-quality internet data, with goal-conditioned offline RL and language models for teaching dialogues as examples.

Your comment is kept while you sign up.Sign in
Levine reads deep RL three ways, from simulator inversion to robots that practise on their failures.
Deep reinforcement learning faces open problems that Sergey Levine frames through three readings of the field in UC Berkeley's CS 285. The first treats reinforcement learning as an engineering tool that inverts a physical simulator into a control law. Moravec's paradox, set against Deep Blue and AlphaGo, then motivates agents that learn in the physical world. Reset-free multi-task dexterous manipulation turns failures into practice, and Parrot's behavioral priors bootstrap exploration. The third reading casts reinforcement learning as a framework for decisions from low-quality internet data, with goal-conditioned offline RL and language models for teaching dialogues as examples.
Deep reinforcement learning from human preferences
Deep Dynamics Models for Learning Dexterous Manipulation
Reset-Free Reinforcement Learning via Multi-Task Learning: Leveraging Dexterous Manipulation Behaviors Without Human Intervention
Parrot: Data-driven behavioral priors for reinforcement learning
Actionable Models: Unsupervised Offline Reinforcement Learning of Robotic Skills
Zero-Shot Goal-Directed Dialogue via RL on Imagined Conversations
This lecture
Lecture 71 of 73 of Deep Reinforcement Learning.
Choose a node to see how it connects to this lecture. What comes before is on the leftat the top. What comes after is on the rightat the bottom. The ring is what it teaches. People and places are above. Other lectures and books are below.
Press a chosen node again to open it.