# SafeDojo: Safe Reinforcement Learning for VLA via Interactive World Models

- Source: arXiv cs.RO
- Published: 2026-10-22T04:00:00.000Z
- RoboHot: https://robohot.news/en/items/yfg6q0j6m45uzsc4nbijt2l9b
- Original: https://arxiv.org/abs/2606.20698

## Summary

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.