Energy-efficient trajectory planning for a class of industrial robots using parallel deep reinforcement learning

With the widespread application of industrial robots, there is growing interest in optimizing their energy consumption during motion processes. Traditional methods typically rely on nonlinear optimization, which can effectively reduce the energy consumption for a given task, but the optimization is time-consuming so that challenging for fixed continuous production lines with evolving tasks. Therefore, for a class of industrial robots with elbow structure, this paper proposes a novel real-time energy-efficient trajectory generation method based on parallel Deep Reinforcement Learning (DRL). The proposed method trains a trajectory generator adaptable to different tasks within a fixed scenario using the Deep Deterministic Policy Gradient algorithm in a simulated environment. During actual operation, the generator rapidly produces trajectories for the real robot to execute. To ensure trajectories meet task requirements and dynamic constraints, this paper designs heuristic rewards for DRL agent. In training process, we enhance training efficiency by simplifying the dynamic model of the class of robots being studied into several sub-modules for parallel training, and we improves the DRL networks by introducing an autoencoder. Through experimental comparison, the simplied model maintains high accuracy in energy calculation and this method can reduce the energy consumption of the robot by 23.21%. Compared with the nonlinear optimizer-based methods, the proposed method can significantly reduce trajectory generation time with only a slight decrease in optimization effectiveness.