Safe Vision Language Action Models via Barrier Enhanced Flow Matching

What the math gives to ML

The paper's transferable contribution is to treat safety enforcement as a modification of the continuous generative vector field rather than as a post-hoc projection of the final action. This creates a useful interface for flow-based samplers: a differentiable barrier over an entire predicted action chunk can impose a forward-invariance condition at every solver step while minimally perturbing the learned velocity. The smooth minimum aggregation converts many waypoint and constraint checks into one scalar differentiable condition. The most direct ML experiment is a barrier-corrected flow-matching sampler that applies the smallest velocity adjustment satisfying a control-barrier inequality, measuring safety, task success, and sampling overhead against final-action filtering.