Safe Near-Optimal Reinforcement Learning for Robotic Motion Planning Using High Order Control Barrier Function

Motion planning remains a critical research challenge in robotics, particularly for nonlinear systems where balancing safety, optimality, and computational efficiency poses inherent trade-offs. While numerous planners have been proposed to address these competing objectives, reconciling rigorous safety guarantees with real-time performance remains an open problem. This paper investigates safe near-optimal robotic motion planning and introduces a reinforcement learning (RL) framework integrating a neural network (NN)-approximated target navigator and a high order control barrier function (HOCBF)-based safeguarding controller. Lyapunov analysis provides formal guarantees on the near-optimality of the derived controller, the convergence of the robot to its target state, and the forward invariance of the robot within the collisionfree set. Experiments on a four-dimensional unicycle-like robot validate the theoretical findings.