Microsoft's ShieldVLA Cuts VLA Safety Costs by 57% Using HJ Reachability

MicrosoftResearch · hf · 2026-09-22

Microsoft Research proposes ShieldVLA, a safety-aligned fine-tuning framework for Vision-Language-Action (VLA) models.

Problem: existing methods rely on Lagrangian soft penalties on expected cumulative cost, causing residual violations or overly conservative behavior; dense per-step safety annotations are also scarce in visual domains.
Method: builds on Hamilton-Jacobi (HJ) reachability, learning a model-free approximation of the reachability value function directly from visual observations to estimate the safe operating region. A learned safety critic gates policy optimization, separating reward maximization in feasible regions from recovery near unsafe states, avoiding persistent reward-cost trade-offs.
Scalable supervision: rubric-based VLM safety scores convert semantic feedback into structured critic targets without manual cost labels.
Results: across five navigation and manipulation benchmarks and multiple VLA backbones, ShieldVLA cuts cumulative safety cost by 57% on average and improves task success rate by +0.13 over SafeVLA.

More from Embodied