Dynamically Feasible Planning and Control in Complex Environments: a Scalable Systematic Approach

What the math gives to ML

The paper provides a constructive reference-governor mechanism for enforcing pointwise safety in stable discrete-time linear dynamics under non-convex constraints represented as unions of polytopes. Its transferable asset is an offline-computed safe-command set combined with a cheap online interpolation that moves a commanded reference toward a target only when the resulting trajectory remains safe. A direct neural-network use is to treat an SSM or recurrent hidden state as the controlled linear subsystem and insert a governor between a raw latent target and the state-update command. This gives a falsifiable safety-versus-convergence tradeoff: exact-model trajectories should remain inside the prescribed polytope union, while commands to strictly admissible targets should converge in finite time.