The Robot Learned Where Reward Had to Stop

Safety fine-tuning for vision-language-action models often treats violations as a soft penalty, forcing one objective to trade reward against risk. ShieldVLA instead learns a model-free approximation of a Hamilton–Jacobi reachability value function from visual observations. The critic separates ordinary reward optimization inside the feasible region from recovery behavior near unsafe states; rubric-based vision-language scores provide training targets without dense manual cost labels. Across five navigation and manipulation benchmarks and multiple VLA backbones, the authors report 57% lower cumulative safety cost on average and a 0.13 gain in task success over SafeVLA. Those benchmark results do not constitute a formal guarantee for an untested physical deployment.