Converting Biomechanical Models from OpenSim to MuJoCo

Abstract

OpenSim is a widely used biomechanics simulator with several anatomically accurate human musculo-skeletal models. While OpenSim provides useful tools to analyse human movement, it is not fast enough to be routinely used for emerging research directions, e.g., learning and simulating motor control through deep neural networks and Reinforcement Learning (RL). We propose a framework for converting OpenSim models to MuJoCo, the de facto simulator in machine learning research, which itself lacks accurate musculo-skeletal human models. We show that with a few simple approximations of anatomical details, an OpenSim model can be automatically converted to a MuJoCo version that runs up to 600 times faster. We also demonstrate an approach to computationally optimize MuJoCo model parameters so that forward simulations of both simulators produce similar results.

Breathing Life Into Biomechanical User Models

“…This introduces a challenge for biomechanical simulation software, which typically has not been designed for such use cases. In our work, we convert models validated by biomechanics researchers into a faster simulator [28]. An alternative would be to use a simplifed simulation model and apply machine learning to predict the omitted details such as state-dependent joint actuation torque limits and muscle-based energy expenditure [30]; however, this requires generating training data using a realistic simulator, and the learned prediction model is inherently less accurate and general than the simulator itself.…”

“…Unity (Figure 2). The implementation can be extended with additional biomechanical models, for instance, by converting them from OpenSim [28], perception models, and interactive tasks. The user model learns to interact with the environment through a taskspecifc reward function.…”

DEP-RL: Embodied Exploration for Reinforcement Learning in Overactuated and Musculoskeletal Systems

“…The muscle also has low-pass filter characteristics, making the control problem hard for classical approaches-in addition to the typical redundancy of having more muscles than DoF. In all our experiments, we use the MuJoCo internal muscle model, which approximates these characteristics and has been used for other muscle-based studies [17,11,23,24]. One limitation consists in the non-elasticity of the tendon.…”

“…All tasks except OstrichRL [17] were constructed from existing geometrical models in MuJoCo [23,36] from which we created RL environments. We additionally created variants of ostrich-run involving perturbations, i.e.…”

Reinforcement Learning Control of a Biomechanical Model of the Upper Extremity

“…In particular, computing dynamic actuator lengths (which significantly affect the forces produced by muscle activation patterns) has still proven challenging in MuJoCo[20].November 17, 2020 10/20…”