Research of Improved TD3 Robotic Arm Path Planning using Evolutionary Algorithm

Nowadays, application of automated intelligent robot arm devices to improve industrial production efficiency has become a popular research field in the world. The previous off-line path planning method of robotic arm has the inadequacies of low efficiency and slow speed. Although the deep reinforcement learning has accomplished many achievements in the path planning of control manipulator, there are still some problems such as long training time and low planning accuracy. To solve the abovementioned issues, we propose an improved Twin Delayed Deep Deterministic policy gradient (TD3) algorithm (Improved Cross-Entropy Method-TD3: ICEM-TD3) for the path planning of the robotic arm. First, this paper combines evolutionary strategies with TD3 to generate action networks. Then the exploration of TD3 algorithm in the action space is replaced by the exploration in the parameter space. In addition, this paper designs a new reward function to weaken the redundancy of planning and accelerate the convergence speed of training. Finally, the Gazebo simulator is adopted to verify the proposed algorithm, and the results illustrate that the proposed algorithm can greatly improve the accuracy of the path planning of the manipulator using deep reinforcement learning.