Abstract:

In this paper, we propose an adaptive dynamic programming (ADP) framework augmented with a learned control barrier function (CBF)–based safety filter for continuous-time linear systems with unknown dynamics in safety-critical scenarios. Using adaptive dynamic programming, a sub-optimal feedback controller is learned from input–state data. The unknown terms in the CBF-based safety constraints are approximated from online data using neural networks, and the resulting learned CBF constraints are incorporated into a quadratic program-based safety filter that modifies the control input to ensure satisfaction of the safety constraints. The effectiveness of the proposed control methodology is illustrated via an obstacle-avoidance example.