Event
SafeDojo Uses World-Model Imagination to Improve VLA Policy Safety
1 reports · 1 sources
Kai Tang et al. propose SafeDojo, a model-based safe reinforcement learning framework: it performs online reinforcement learning on top of an interactive video world model, uses a ResNet classifier to score predicted future frames, incorporates a lightweight safety head to predict per-step contact probability, and then employs Lagrangian-constrained GRPO to balance task execution and trajectory safety, thereby improving vision-language-action (VLA) policies and avoiding costly real-world exploration or hand-designed safety functions. On SafeLIBERO Level I, SafeDojo improves the average safe success rate by 8.25 percentage points over the strongest baseline; in the zero-shot Level II setting, its performance is comparable to the strongest model-free baseline. In deployments on a real Franka robot across five tasks, SafeDojo achieves the best average task success rate and safe success rate.
Report timeline
- SafeDojo: Safe Reinforcement Learning for VLA via Interactive World Models
SafeDojo proposes a model-based safe reinforcement learning framework that improves vision-language-action (VLA) policies through world-model imagination, avoiding costly real-world exploration or reliance on hand-crafted safety functions. The method performs online reinforcement learning on top of an interactive video world model, using a ResNet success classifier to score predicted future frames, paired with a lightweight safety head that predicts per-step contact probabilities from the current latent context and candidate action chunks, then balances task execution and trajectory safety signals through Lagrangian-constrained GRPO.
arXiv cs.RO ·