Event
Levine says supervised extraction conditioned on advantage can achieve RL optimum
2 reports · 2 sources
Sergey Levine posted that extracting a policy using only supervised learning conditioned on advantage can achieve state-of-the-art results in reinforcement learning (RL), provided several important details are handled correctly. Levine did not explain in the post what these details specifically are, nor did he provide corresponding experimental data or a paper link.
Developments
- VLA success rate SOTA confirmed at 80% · 1 reports
- Advantage-Conditioned Supervised Extraction Achieves Optimal RL · 1 reports
Report timeline
- VLA success rate SOTA confirmed at 80%
Just want to confirm: the SOTA for VLA is an 80% success rate, right?
Ken Goldberg ·
- Advantage-Conditioned Supervised Extraction Achieves Optimal RL
It turns out that getting a few important details right is all it takes: extracting policies using only advantage-based supervised learning can achieve state-of-the-art results in reinforcement learning (RL)!
Sergey Levine ·