Reproduce “RoboLongR: Causally Consistent Rewards Teach Robots to Progressively Master Long-Horizon Manipulation” 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
Zixuan Chen, Lin Cao, Fenggang Zhang, Haotong Zhang, Yangtao Chen, Jieqi Shi, Jing Huo, Yang Gao
RoboLongR trains history-sensitive rewards for long-horizon robot tasks by learning from reordered demonstration segments and recurrent memory. It then uses these rewards to guide policy improvement. On 18 held-out RLBench tasks, it reaches 31.8% average complete-task success after 20 demonstrations per task and online interaction, versus 22.6% for ReWiND. The result measures adaptation, not zero-shot performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.