Reproduce “EmbodiedSmith: Scaling Embodied Data through Recursive Self-Improvement Flywheel in Simulation” 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
Yikai Qin, Yifei Deng, Mingjian Liang, Wenxuan Song, Zepeng Lin, Zhiyi Jiang, Jiajun Fu, Qiao Sun and 8 more
EmbodiedSmith links scene and robot-task generation through execution feedback: tasks and scenes are revised after simulated rollouts fail. The system plans tasks around scene objects and robot skills, then repairs plans or makes targeted scene edits and revalidates them. In Fixed-Long, feedback raises average success rate from 0.0% to 71.6%. Mixed training data improves randomized evaluation but lowers fixed-layout performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.