Simultaneous Learning and Planning using Rapidly Exploring Random Tree* and Reinforcement Learning

The paper proposes an approach to learn and plan simultaneously in a partially known environment. The proposed framework exploits the Voronoi bias property of Rapidly exploring Random Tree* (RRT*), which balances the exploration-exploitation in Reinforcement Learning (RL). RL is employed to learn policy (sequence of actions), while RRT* is planning simultaneously. Once policy is learned for a fixed start and goal, repeated planing for identical start and goal can be avoided. In case of any environmental uncertainties, RL dynamically adapts the learned policy with the help of RRT*. Apparently, both learning and planning complement each other to handle environmental uncertainties dynamically in real-time and online. Interestingly, more the proposed algorithm runs, its efficiency increases in terms of planning time and uncertainty handling capability over the contender algorithm (i.e., RRT*). Simulation results are shown to demonstrate the efficacy of the proposed approach.