Deep Reinforcement Learning-based Multi-task Optimization Algorithm for Industrial Robots

With the advancement of intelligent technologies, robots are increasingly integrated into various aspects of human production and daily life. As a result, the acquisition of operational skills by robots has emerged as a prominent research focus. However, due to the growing complexity of tasks, conventional machine learning methods are no longer sufficient to address the learning demands of robotic manipulation. Deep reinforcement learning (DRL) realizes autonomous learning and decision making of robot operation skills by controlling the policy network to interact with the environment. However, in practice, the high cost of robot interaction and the low efficiency of reinforcement learning samples have become bottlenecks restricting the wide application of DRL. In order to improve the robot's decision-making ability for complex tasks, this paper adopts a Dense2Sparse by network resetting (Dense2Sparse) to solve the problem of the difficulty in designing traditional dense reward functions. At the same time, a pruning mechanism is introduced to address the issue of slow convergence in multitask deep reinforcement learning models. The proposed approach is validated through a series of robotic arm grasping tasks.