Reinforcement Learning-based DDPG and PRM Integration for Optimal Navigation in Varying Complex Environments

We have Autonomous navigation in various environments is critical for real-world applications.Adaptation to new environments and complexities is crucial for autonomous mobile robots.This project investigates feasibility of finding the optimal sample time and average reward structure for a given size of environment using Reinforcement Learning based Deep Deterministic Policy Gradient (DDPG) technique.This project finds the optimal connection distance and number of nodes to be given in finding the feasible path using Probabilistic Roadmap (PRM).Integrating these two techniques will enable a mobile robot to avoid obstacles while reaching the destination by following an optimal path.As a case study a mobile robot is trained using DDPG in three different environments of size 26x27 units with varying complexity.By varying sample time and reward, different agents were created.Analysis is done based on the obstacle avoidance and maximum distance travelled by mobile robot.By varying number of nodes and connection distance feasible paths found using PRM.By integrating these two DDPG and PRM techniques, the mobile robot path length to reach the destination is reduced by 40%.