Online Learning Based Control for Mobile Vehicle to Avoid Static Obstacles via Kernel Based Safe Dynamic Model

This paper addresses the path planning for mobile vehicle subject to state constraint i.e., avoiding static obstacles by finding the kernel based safe dynamic model by using the Gaussian kernel via online sampling and without knowing the dynamic of the system to extract the benefit of model free reinforcement learning. A proper reward function is designed to find the collision free optimal path at each time step. A proper barrier function which is function of distance in term of feature of controller to avoid collision with obstacle is augmented in reward function. At each time step, safe optimal control policy is find locally by generating safe samples around the vicinity of the current state to solve the curse of dimensionality issue and then, solving Bellman optimality equation like Q-learning algorithm that used the safe kernel based dynamic model, hence combine the effect of model based reinforcement learning at this point. Finally, simulation results demonstrate the validity of the proposed algorithm.