A Trajectory Planning Method for Automatic Water Fetching by a Robotic Arm Based on Deep Reinforcement Learning

The study focuses on trajectory planning for an assistive robotic arm to autonomously fetch water, using deep reinforcement learning (DRL) algorithms to assist patients in complex environments. In real-world scenarios, the robot base is responsible for reaching a designated target location, while the robotic arm performs tasks of grasping and placing objects. Given the different precision requirements for controlling each, it is evident that the robotic arm requires continuous, precise trajectory control. Based on this, considering the motion differences between the robotic arm and its base, a combination of sparse and continuous reward strategies is employed to control the movement trajectory of the assistive robotic arm. A novel reward function is designed to aid the arm in better exploring the environment. The approach integrates the Actor-Critic algorithm with the Deep Deterministic Policy Gradient (DDPG) algorithm to train the assistive robotic arm. Ultimately, a multi-task, multi-objective assistive robotic arm system is trained using DRL, achieving the task of fetching, filling, and delivering a cup of water in a complex environment.