D-SafeMPC: Diffusion-Driven Safe Model Predictive Control with Discrete-Time Control Barrier Functions

What the math gives to ML

The paper provides a concrete safety interface between a learned trajectory generator and nonlinear control: represent collision avoidance with state-dependent barrier functions, augment the state with obstacle dynamics, and enforce safety during every refinement step rather than only after sampling. The transferable asset is the differentiable construction of a worst-case safety margin from multiple constraints, combined with dynamics-aware correction of a proposed trajectory. A practical neural-network adaptation is to guide diffusion trajectory sampling using discrete-time control-barrier penalties, so the denoiser moves toward dynamically feasible safe regions instead of relying on rejection sampling. This is most useful for diffusion policies, model-based planners, and sequence models producing actions under changing obstacles.