Learning Input-Constrained Funnel Controllers from State Trajectory Data

What the math gives to ML

The paper offers a transferable mechanism for learning a prescribed-performance funnel and a feedback controller from state trajectories without observing expert actions, while explicitly accounting for hard actuator limits. Its key asset is the combination of data-derived transient envelopes, state-dependent feedback gains, and feasibility-driven active-set synthesis, backed by conservative actuator-authority and local trajectory-density certificates. A neural implementation should use demonstrations to construct a time-varying error envelope, train a bounded state-feedback network inside that envelope, and reject or reduce gains that cannot satisfy actuator authority. The resulting method makes falsifiable predictions: trajectories should remain inside the learned funnel, and the maximum feasible feedback gain should scale inversely with the funnel radius and actuator limits.