Simulation study of exoskeleton-musculoskeletal coupling for exoskeleton assistance via multi-agent reinforcement learning.

To address the wearable exoskeletons' core bottlenecks, difficulty in personalized adaptation, over-reliance on iterative human experiments, and poor generalization across multiple working conditions. This study proposes a musculoskeletal-exoskeleton coupled assistance framework based on heterogeneous multi-agent reinforcement learning (MARL), with deep integration of biomechanical principles. A high-fidelity OpenSim simulation environment is constructed, incorporating a bilaterally symmetric 7 degrees of freedom (7-DOFs) lower-limb musculoskeletal model and a hip-assist exoskeleton, which are modeled as heterogeneous agents and trained under the Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (MATD3) algorithm's "centralized training and distributed execution" (CTDE) paradigm. Desired kinematic trajectories from healthy human gait data are embedded as biomechanical priors; a multi-objective reward function constrains joint tracking errors, posture instability, core muscle activation (aligning with "core-auxiliary muscle" load division law), and abrupt exoskeleton actions, while an Actor network action smoothness term suppresses control fluctuations to match human musculoskeletal dynamic response characteristics. Simulation results (0.86 m/s, 1.2 m/s and 1.5 m/s) demonstrate high-precision symmetric hip/knee kinematic tracking; the iliopsoas peak force reduced from > 3000 N to < 1600 N; and by 10-25% lower ground reaction force (GRF) bimodal peaks. This work verifies that biomechanics-integrated heterogeneous MARL can learn robust, speed-adaptive assistance policies in simulation, offering a scalable pathway to accelerate exoskeleton controller development and address long-standing industry challenges.