Reproduce “When the Simulator Fails, Not the Policy: Auditing Contact Artifacts in Simulated Robot Benchmarks” 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
Alex Zhou, Junhao Wu, Xueyang Zhou, Yixiao Huo, Hongcheng Gao, Di Zhu, Zhenghao Zhang, Yongchao Chen
The authors introduce SimuGuard, which links suspicious simulator contacts to episode outcomes through per-step recording, exact replay, and controlled reruns. In RoboTwin, they trace false contacts to a collision-detection threshold affecting centimeter-scale geometry in PhysX 5.3.1 or earlier. On five container-placement tasks, LingBot-VA passes 89/100 episodes in the released engine and 98/100 after repair; this comparison is not statistically conclusive.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.