Reinforcement Learning Based Automated Path Planning in Garden Environment using Depth - RAPiG-D

Path planning by employing Reinforcement Learning is a versatile implementation that can account for the ability of a robot to autonomously map any unknown environment. In this paper, such a hardware implementation is proposed and tested by making use of the SARSA algorithm for path planning and by utilizing stereovision for depth estimation based obstacle detection. The robot is tested in a cell-based environment – 3x3 with 2 obstacles. The goal is to map the environment by detecting and mapping the obstacles and finding the ideal route to the destination. The robot starts at one end of the environment and runs through it for a specified number of episodes, and it is observed that the robot can accurately identify and map obstacles and find the shortest path to the destination in under 10 episodes. Currently, the destination is a fixed point and is taken as the other diagonal end of the environment.