Neuro-Symbolic Safety Guidance for Vision-Language-Action Models via Constrained Flow Matching

What the math gives to ML

The paper's transferable contribution is to treat safety guidance as a projection of an entire predicted action trajectory, rather than checking only the next action. During Euler integration of a flow-matching sampler, a differentiable collision surrogate can produce linearized inequality constraints, and a minimum-norm correction can project the denoising update back into the locally safe set. This is directly applicable to diffusion or flow-based trajectory generators, especially when long-horizon errors compound and an inference-time correction is cheaper than retraining the policy.