Deep Reinforcement Learning with Inverse Jacobian based Model-Free Path Planning for Deburring in Complex Industrial Environment

In this study, we present an innovative approach to robotic deburring path planning by combining deep reinforcement learning (DRL) with an inverse Jacobian strategy. Existing model-based path planning methods, including sampling-based approaches, often suffer from computational complexity and challenges in capturing the dynamics of deburring systems. To overcome these limitations, our novel DRL-based framework for path planning leverages experiential learning to identify optimal deburring trajectories without relying on predefined models. This model-free approach is particularly suited for complex deburring scenarios with unknown system dynamics. Additionally, we employ an inverse Jacobian technique with a time-varying gain module (eta(t) = e<^>2t) during training, which yields remarkable benefits in terms of exploration-exploitation balance and collision avoidance, enhancing the overall performance of the DRL agent. Through a series of experiments conducted in a simulated environment, we evaluate the efficacy of our proposed algorithm for deburring path planning. Our modified DRL-based approach, utilizing inverse kinematics with a time-varying gain module, demonstrates superior performance in terms of convergence speed, optimality, and robustness when compared to conventional path planning methods. Notably, in comparison to algorithms like sampling-based strategies, our model-free DRL-based approach outperforms these methods, achieving an exceptional average success rate of 97%. The integration of the inverse Jacobian technique further enhances the effectiveness of our algorithm by effectively reducing the state space dimensionality, leading to improved learning efficiency and the generation of optimal deburring trajectories.