Summary

Existing VLA safety measures only prevent collisions caused by the robot’s immediately next action, which is insufficient for flow-matching VLAs that determine actions by predicting a full trajectory through iterative neural flow matching. This paper introduces a neuro-symbolic safety guidance mechanism that integrates symbolic safety constraints directly into the flow matching process, enabling predictive collision avoidance across the entire trajectory horizon.

Key Contributions

  • Neuro-symbolic safety guidance embedded within the iterative flow matching inference process
  • Predictive collision avoidance over the entire action trajectory, not just the next action
  • Compatible with existing flow-matching VLA backbones without retraining
  • Formal safety guarantees through constrained optimization during inference

Significance

The first safety framework designed specifically for the trajectory-level inference of flow-matching VLAs, enabling safe deployment in environments with dynamic obstacles without compromising task performance.