Reinforcement Learning Based Musculoskeletal Finger Tip Multi-Objective Motion Control and Generalization

Musculoskeletal robots, which mimic the arrangement and actuation of human or other organism muscles and bones, offer significant advantages including high compliance, robustness, and dexterity. However, their complex, redundant, and nonlinear structures present substantial challenges in control. Recent advancements in musculoskeletal robots have highlighted difficulties in executing and generalizing motion tasks, making this a key focus in current research on motion control. In this paper, we address the challenge by formulating a multi-objective motion control and generalization task for a musculoskeletal finger robot. We tested various control algorithms using three widely-adopted reinforcement learning methods, each with simple reward settings. Statistical errors from these algorithms were compared across the multi-objective motion and generalization tasks. Furthermore, our comparative analysis of the movement processes with motor-driven robots demonstrates that musculoskeletal robots exhibit superior compliance and stability during motion.