Research map / Reinforcement learning / Policy gradient and actor-critic
OGPO: Sample Efficient Full-Finetuning of Generative Control Policies
Sarvesh Patil, Mitsuhiko Nakamoto, Manan Agarwal, Shashwat Saxena, Jesse Zhang, Giri Anantharaman, Cleah Winston, Chaoyi Pan, Douglas Chen, Nai-Chieh Huang, Zeynep Temel, Oliver Kroemer and 5 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add success-buffer behavior cloning regularization (ablation) | OGPO · Robomimic state-based tasks | success rate | Improved |
| add Best-of-N inference with critic (ablation) | OGPO+ · Robomimic Toolhang | success rate | No clear effect |
| use conservative advantages (all ensemble critics agree on sign) (ablation) | OGPO+ · Robomimic image-based tasks | success rate | Improved |
| use conservative advantages (ablation) | OGPO+ · offline-to-online adaptation | performance dip | Improved |
| add Q-variance reduction of TD targets (ablation) | OGPO · pixel-based Robomimic | critic accuracy | Mixed |
| use PPO policy extraction instead of AWR (flow-matching loss weighting) (ablation) | OGPO · Robomimic tasks | success rate | Improved |
| use PPO likelihood ratio instead of exponentiated advantage weights (ablation) | OGPO · Robomimic tasks | success rate | Improved |
| backpropagate through denoising chain (BPTT) instead of zero-order PPO (ablation) | OGPO · Robomimic tasks | success rate | Hurt |
| remove negative-advantage gradients (ablation) | OGPO · Robomimic Transport | success rate | Hurt |
| normalize advantages by group standard deviation (GRPO-style) (ablation) | OGPO · Robomimic | success rate | Hurt |
| add CalQL+WSRL offline-to-online recipe (ablation) | OGPO · Robomimic | success rate | No clear effect |
| use pixel-based critic with vanilla OGPO (ablation) | OGPO · Robomimic Square | success rate | Hurt |
Closest papers by meaning
- GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.91)
- Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.90)
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.89)
- OMPO: A Unified Framework for RL under Policy and Dynamics Shifts (ICML 2024 · similarity 0.89)
- Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.89)
- Aligning Flow Map Policies with Optimal $Q$-Guidance (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
