A Multi-Policy Framework for Manipulator Trajectory Tracking Based on Feature Extraction and RL Compensation

Trajectory tracking for manipulators, especially when faced with complex trajectories, remains a formidable challenge in robotics. This paper introduces an approach to address this challenge with the framework we propose called Cluster-SAC. Our strategy hinges on the integration of feature extraction techniques to segment complex trajectories, assigning each segment to be adeptly tracked by distinct SAC agents. Through this approach, the precision of trajectory tracking is optimized, as each segment benefits from a dedicated learning agent. Validated with the 7-DoF manipulator of Franka Emika Panda Robot in the Gazebo simulation, results showcase faster training convergence and superior tracking accuracy on arbitrary given trajectories.