Reinforcement Learning-Based Trajectory Planning for Fracture Reposition Robot: A Comparative Study with Cartesian Methods

This study aims to address the limitations of traditional fracture reposition trajectory planning methods-namely, poor adaptability, low efficiency, and difficulty handling multi-objective constraints-and investigates the potential of a reinforcement learning (RL)-based approach for optimization. A simulation platform for a parallel fracture reposition robot was developed, integrating a spring-mass system for musculoskeletal modeling and force estimation. The RL framework defined the state space, action space, boundary conditions, and reward functions to enable autonomous trajectory planning. Comparative experiments against Cartesian linear planning demonstrated the proposed method's advantages. The RL agent achieved an average episode duration of 28.68 seconds, a cumulative reward of 166.42, and completed repositioning tasks without re-planning, indicating high adaptability and efficiency. Compared to Cartesian planning, the method reduced peak musculoskeletal force by 11.68 %, average force by 15.89 %, and improved repositioning efficiency by 13.79 %. These results validate the approach's effectiveness in generating smooth and precise trajectories under complex, multi-objective conditions.