Reproduce “ReTrace: Cross-Stage Failure Attribution for Self-Improving Robotic 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
Yue Yu, Bin Zhu, Junhui Li, Jiayu Wang, Shijie Zhou, Jingjing Chen
ReTrace links later manipulation failures to earlier decisions, tests focused corrections in attempts, and saves supported practices as reusable skills without changing model parameters. On a Franka Research 3, GPT-6 Astra-low achieved 80% success on plug insertion and 100% on ball pouring with skill, versus 0% without skill in 10 trials each; evidence is limited to four tasks and tested configurations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.