Reproduce “SafeDojo: Safe Reinforcement Learning for VLA via Interactive World Model” 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
Kai Tang, Peidong Jia, Zhong Chu, Jiajun Cao, Jixian Wu, Bohua Zhang, Rui Ma, Fangyuan Zhao and 11 more
SafeDojo optimizes vision-language-action policies in an action-conditioned video world model, where imagined futures receive separate task-success and predicted-contact scores before a constrained policy update. On SafeLIBERO Level I, it reports 53.25% safe success, versus 45.00% for SafeVLA. In Level II, the rates are 49.62% and 49.50%, respectively; its safety budget uses predicted costs, not a verified real-world guarantee.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.