Reproduce “RoboMD: Uncovering Robot Vulnerabilities through Semantic Potential Fields” 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

Som Sagar, Jiafei Duan, Sreevishakh Vasudevan, Yifan Zhou, Heni Ben Amor, Dieter Fox, Ransalu Senanayake

The ROBOMD framework introduces a deep reinforcement learning approach operating within a continuous vision-language semantic embedding space to proactively diagnose vulnerabilities in pre-trained robot manipulation policies. This method effectively identifies and ranks failure modes across diverse environments, improving policy robustness by 7.35% (from 85.48% to 92.83%) with 85% less data for fine-tuning.

Key insights

  • Transforming robot vulnerability diagnosis into an active search within a continuous semantic embedding space enables systematic, scalable, and safe identification of failure modes, including previously unknown ones.
  • Multimodal inputs (image and text) combined with contrastive loss create an effective embedding space where success and failure outcomes form distinct, continuous clusters, serving as a reliable 'potential field' for RL-driven exploration.
  • The use of PPO within this semantic space, supported by theoretical guarantees, enables sample-efficient exploration of the success/failure boundary and accelerates convergence in identifying critical policy weaknesses.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.