Reproduce “StructRL: Online Structured Reinforcement Learning for Long-Horizon Vision-Language-Action Tasks” 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
Ziyi Yin, Sangmin Woo, Kang Zhou, Sungyeon Kim, Aosong Feng, Haibo Ding, Jun Huan
StructRL trains vision-language-action policies for long-horizon robot tasks using simulator-checkable milestones, ordered prerequisites, and demonstration-based pacing, rather than only final success. An LLM selects and organizes verifiable subtasks before reinforcement learning, while eligible steps receive intermediate credit. On RoboCasa365 Composite-Seen, macro success rate, the mean of task-level rates, is 49.1% for GR00T-N1.5 versus 41.5% for SimpleVLA-RL, in simulation only.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.