QF3 project page live: interactive explainer for fast flow RL on robots
andrew_n_carr · x · 2026-10-13
Project page and paper (arXiv 2610.08789) for QF3 are up, from UC Berkeley and Amazon FAR with authors including Ken Goldberg, Pieter Abbeel, and Angjoo Kanazawa.
QF3 stabilizes off-policy RL for flow policies by clipping critic-driven velocity updates around replay data and masking gradients where the critic can't be trusted, preventing the policy from chasing overestimated rewards. Validated on humanoid locomotion and manipulation fine-tuning, with an interactive explainer and demo videos on the project page.
Related event: UC Berkeley's QF3 Enables Stable RL Fine-Tuning of Flow Policies(2 posts)→
More from Embodied
- Dev ships 'EveryDay Carry' firmware for AI Passport, hailed as best network radio — ezshine · 2026-10-13
- RoboARCade launches: agents directly control robots, best GPT-6 setup tops out at 49.8% success — leto__jean · 2026-10-13
- Free-Init: scan-free, motion-free initialization for Doppler LiDAR-inertial systems — rsasaki0109 · 2026-10-13
- On-device robot practices soccer throw-ins with no external tracking — chris_j_paxton · 2026-10-13
- GPT-6 Astra and Claude Opus 5 nearly tied on embodied task: 34-6 over 40 votes, 16-point Elo gap — qinzytech · 2026-10-13
- Custom ComfyUI node repurposes Mixamo's 3D rigs to pose 2D characters — Willonfire8 · 2026-10-13