SafeDojo teaches robots safe actions via imagined world models

Safe control is a critical prerequisite for deploying embodied AI in real-world environments, but existing safe reinforcement learning (RL) methods often require costly real-world exploration or rely on hand-crafted safety functions that don't scale to complex vision-language-action (VLA) models. To address this, researchers from multiple institutions developed SafeDojo, the first model-based safe RL framework designed specifically for VLA policies. The key innovation is an interactive video world model that allows the agent to imagine future scenarios and learn safe actions without physical trial and error.

SafeDojo performs online RL on top of this world model, which generates action-conditioned future predictions. A tailored ResNet success classifier estimates per-step task progress from imagined frames, while a lightweight safety head predicts per-step safety costs from latent context and the proposed action chunk. These decoupled signals are balanced through a Lagrangian-based constrained GRPO objective, enabling coordinated improvement in both task success and safety. On the SafeLIBERO benchmark, SafeDojo outperformed all baselines—including inference-time safety, model-free RL, and model-based RL—achieving the best average safe-success rate on both levels and an 8.25 percentage-point improvement on Level I. Real-world deployment on a Franka robot arm across five tasks further demonstrated the best average task and safe-success rates. The results position world model-based safe RL as a scalable and generalizable path toward safe embodied intelligence.