Multipath Planning Acceleration Method With Double Deep R-Learning Based on a Genetic Algorithm

摘要

Autonomous navigation is a well-studied field in robotics requiring high standards of efficiency and reliability. Many studies focus on applying AI techniques to obtain a high-quality map, a precise localization, or improve the proposed trajectory to be followed by the agent. As traditional planning methods need a high-quality map to obtain optimal trajectories, this paper addresses the problem of multipath map-less planning, and proposes a novel multipath planning algorithm (Double Deep Reinforcement Learning - Enhanced Genetic (DDRL-EG)) for mobile robots in an unknown environment. It combines Double Deep Reinforcement Learning (DDRL) with Heuristic Knowledge (HK), Experience Replay (ER), Genetic Algorithm (GA), and Dynamic Programming (DP), allowing the agent to reach its target successfully without maps. In addition, it optimizes the training time and the chosen path in terms of time and distance to the target. A hybrid method is also used in which Semi-Uniform Distributed Exploration (SUDE) is employed to determine the probability that the action is decided based on directed knowledge, hybrid knowledge, or autonomous knowledge. The performance of DDRL-EG is compared with two other algorithms in two different environments. The results show that DDRL-EG is a more robust and powerful algorithm since with less training, it can provide much smoother and shorter trajectories to the target.