TalkRL: The Reinforcement Learning Podcast Neurips 2024 RL meetup Hot takes: What sucks about RL?
Dec 23, 2024
AI Snips
Chapters
Transcript
Episode notes
Human-in-the-Loop RL
- Incorporate human feedback and domain expertise.
- Avoid relying solely on sparse rewards, which can be ineffective.
Avoid Tabula Rasa
- Don't start from scratch (Tabula Rasa) in RL.
- Pre-train agents with existing knowledge or structure.
Sim-to-Real Gap in Value Functions
- Sim-to-real transfer in value function learning is a significant challenge.
- Learned value functions in simulation often don't translate well to real-world robots.