Optimal Control and Reinforcement Learning for Constrained Mobile Robot Navigation

The research addresses the problem of autonomous navigation of a differentially-driven mobile robot in environments with static obstacles and kinematic constraints. A model-based actor-critic method within the approximate dynamic programming framework is proposed, in which the critic loss is augmented with a canonical equation term derived from the connection between Pontryagin's Maximum Principle and the Hamilton-Jacobi-Bellman equation. Environmental constraints are handled via a soft barrier penalty based on an artificial potential field, integrated directly into the utility function. Experiments on the TurtleBot3 platform demonstrate that the proposed method achieves reliable goal-reaching under obstacle avoidance constraints.