Efficient Constraint-Based Musculotendon Simulation for Biomechanically Accurate Characters

Overview This project will create a next-generation, open-source physics simulator that enables accurate and stable modeling of musculotendon dynamics for graphics and beyond, using a novel constraint-based approach. Unlike traditional methods that apply musculotendon forces, this simulator enforces physiologically informed constraints, allowing for fully implicit and unified simulation of muscles, tendons, skeleton, and environment. The first research thrust introduces a fully implicit simulation framework that incorporates nonlinear, velocity-dependent muscle models directly into constraint formulations, enabling stable and accurate simulation of complex musculotendon dynamics. The second thrust focuses on character control, leveraging reinforcement learning to train muscle-driven controllers that learn activation patterns reflective of real human physiology, improving upon joint-actuated models that often yield unnatural or unrealistic motion. The third thrust validates the simulator through comparison with in-vivo biomechanics and physiological data, as well as benchmarking against standard tools to assess both fidelity and computational performance. The final thrust demonstrates broad applicability by enabling physiologically informed animation, injury-aware character control, and movement optimization in high-impact or high-contact tasks. Keywords: physics-based simulation; musculotendon modeling; virtual human. Intellectual Merit This project will advance knowledge by unifying constraint-based simulation with biomechanically accurate musculotendon modeling, enabling new computational and theoretical insights across computer graphics, biomechanics, and human-centered computing. It will address fundamental challenges in simulating muscle dynamics by developing a fully implicit constraint-based framework capable of capturing nonlinear, velocity-dependent musculotendon behaviors. The project will also develop a new class of muscle-driven controllers using deep reinforcement learning techniques that account for physiological constraints such as fatigue, co-activation, and injury. These controllers will provide insight into how biological structure informs control and muscle coordination, with implications for both digital human animation and robotics. In addition, the simulator will be validated using experimental biomechanical data and compared with standard tools, providing a rigorous benchmark for physical fidelity and computational performance. Altogether, the research will produce foundational knowledge and software tools that significantly advance the theory and practice of human motion simulation. Broader Impacts This project will create a new class of computational tools grounded in biomechanical principles and advanced graphics algorithms, enabling physiologically accurate simulation of human motion for a wide range of applications. By integrating validated muscle models into simulation frameworks, the work will support cross-disciplinary research in sports, health, rehabilitation, and robotics—allowing researchers to model complex, real-world movements that were previously impractical. The tools will be released as open-source software, including a pedagogical version, deep reinforcement learning controllers, validation datasets, and applications. Educational impact will be achieved through curriculum integration and the creation of an interactive platform for teaching the interplay between motion, control, and environment. The project aligns with national priorities in scientific discovery, digital health, and personalized diagnosis, and will be combined with outreach initiatives to engage the public through the compelling subject of human motion.