Path planning of robotic arm based on deep reinforcement learning algorithm

Abstract Robotic Arms are used in many fields due to their high accuracy. A robotic arm has the advantage of solving the same tasks as a human arm because of its similar structure and working repetitively and independently of any human. Setting up a robotic arm's task requires the engineer to plan a path for solving the task. Using an approach with reinforcement learning to make the robotic arm learn the path independently has not been feasible due to robotic arm tasks' sparse rewards. A recently successful method of using Reinforcement Learning is the concept of Hindsight Experience Replay. This article extends the difficulty of existing tasks to two new tasks and examines Hindsight Experience Replay's performance on these tasks. In the first experiment, the robotic arm has to move a ball towards a point that is far outside of its reach, similar to golf. In the second experiment, the task is to toss a ball into a box outside its reach, similar to basketball. Results show that vanilla Hindsight Experience Replay performs poorly on these tasks. Further research on solving the tasks with improvements to Hindsight Experience Replay, like Hindsight Goal Generation and Energy‐Based Hindsight Experience Prioritization, is required to solve them further.