Reproduce “PF-RL: Progress Field Reinforcement Learning via Goal-Conditioned Value Geometry for Vision-Language-Action Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yunpeng Qing, Yilun Kong, Sixu Lin, Ming Zhou, Yiming Fei, Shuang Luo, Yixiao Chi, Haoming Gu and 4 more
PF-RL addresses sparse credit assignment in long-horizon manipulation by measuring progress toward a task goal in learned representation space. It projects current and goal observations through a shared head and uses their distance as value. On LIBERO, it reaches 99.1% success versus 97.1% for RLinf-GRPO; this full-pipeline comparison does not isolate geometry, though ablations favor it in the tested setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.