This paper presents a safe control policy search for autonomous J-turn maneuvers inspired by professional car drivers. These drivers have been highly trained to achieve highly aggressive J-turn maneuvers that can be used as potential vehicle safety strategies in emergency situations. We take advantage of this knowledge by integrating an expert’s instruction into reinforcement learning design. The proposed safe reinforcement learning method uses constrained Markov decision processing to find a risk-free control policy for autonomous J-turn maneuver. The proposed algorithms realize the control of continuous action as a sequence of policies with safety constraints. We demonstrate experimentally the development of policy design for the safe J-turn maneuver from simulations to real-world applications. Paper Link
Leave a Reply