PMM: A Parametric Musculoskeletal Model Aligned with Human Body Shape

Abstract

Simulating musculoskeletal models in a physics-based environment enables natural motion that follows human biomechanics. However, constructing musculoskeletal models tailored to diverse body shapes is both time-consuming and labor-intensive. Moreover, achieving robust control over complex, muscle-driven actuation remains a significant challenge. In this paper, we propose a novel shape-parameterized musculoskeletal model that is both anthropometrically and biomechanically plausible. Our approach supports variation in physical attributes—including height, mass, and body shape—as well as muscle–tendon unit parameters. We use the parametric human body model (SMPL) to represent body shape variations and estimate joint locations. From this parameter space, we retarget a template musculoskeletal model to any target shape, adapting both physical properties and muscle-specific parameters. The resulting individualized musculoskeletal models are simulated in a physics-based environment using learned control policies. A hybrid controller combines a stable proportional–derivative servo (PD-servo) servo for the upper body with a Hill-type muscle–tendon model for the lower body. We train these control policies using deep reinforcement learning (DRL) to control diverse locomotion and solve challenging tasks across a wide range of body configurations and environments.