Optimization Study of Dynamic Weight Adjustment Based on Reinforcement Learning for Trajectory Tracking in Sorting Robots

Abstract

Trajectory tracking control for sorting robots in dynamic warehouse environments is challenging due to environmental uncertainty and frequent disturbances. The performance of traditional model predictive control (MPC) heavily relies on manually pre-tuned weight parameters in its cost function. This fixed configuration limits the ability to autonomously balance multiple objectives—such as trajectory tracking, obstacle avoidance, and energy consumption—in dynamic settings, thereby constraining adaptability and robustness.To address this, this paper introduces a hierarchical reinforcement learning framework for online autonomous adjustment of MPC weights. The framework consists of a high-level meta-controller and a low-level MPC executor. The high-level controller dynamically adjusts the MPC weight matrix based on the global environment state, enabling intelligent prioritization of control objectives. Comparative experiments in a high-fidelity Gazebo simulation demonstrate that the proposed method outperforms both fixed-weight MPC and PID controllers in tracking accuracy, task efficiency, safety, and energy consumption. The results validate the effectiveness of the approach and reveal an interpretable, learning-based decision-making mechanism, offering a reliable solution for high-performance robot control in dynamic environments.