MyoSuite -- A contact-rich simulation suite for musculoskeletal motor control
TL;DR MyoSuite converts validated OpenSim models of the finger, elbow and hand into MuJoCo and builds 204 contact-rich RL tasks on them, from posing to key turning, pen twirling and Baoding balls, with sarcopenia, fatigue, tendon-transfer and exoskeleton variants. The converted models match OpenSim muscle forces within 4.1% RMS error while simulating 60x to 4000x faster.
What they do
Muscle-driven control is pull-only, redundant and non-stationary, but RL benchmarks use joint-level control, and musculoskeletal simulators such as OpenSim are slow and offer limited contact support. MyoSuite brings physiologically validated arm and hand models into MuJoCo and defines dexterous tasks on them. In total there are 204 tasks: 9 task families, 2 difficulties, 3 reset conditions and 8 (or 4) non-stationarity combinations.
How it works
- Conversion pipeline. OpenSim models are converted in three steps: geometry transfer, moment-arm matching (simulated annealing over wrapping-surface sites and sizes) and muscle-force matching (differential evolution over force-length parameters).
- Models. MyoFinger (4 DoF, 5 muscle-tendon units), MyoElbow (1 DoF, 3 flexors and 3 extensors) and MyoHand (29 bones, 23 joints, 39 muscle-tendon units). The policy outputs one neural control signal per muscle.
- Non-stationarity. Sarcopenia (maximal isometric force cut by 50%), a dynamic fatigue model, an EIP-to-EPL tendon transfer, and an elbow exoskeleton modeled as an ideal torque actuator.
Results
- Moment-arm RMS difference to OpenSim is 0.044% (elbow) and 0.38% (hand); force RMS error is 2.2% and 4.1% of peak force. MuJoCo runs 60x to 4000x faster, and its cost does not grow exponentially with muscle count as OpenSim’s does.
- NPG baselines are reported up to 5M samples per task; Baoding balls was solved only after more than 70M samples and in 1 of 5 seeds.
- Tearing the opponens pollicis reduces key rotation because the flexor pollicis longus cannot compensate; after tendon transfer a trained policy fails and needs extensive retraining.
- With a 2 kg load, the exoskeleton keeps elbow error near 4.3 degrees while cutting muscle activation by 54-66%.
Limitations
The authors call the models a first approximation needing further development and validation. MuJoCo muscles use infinitely stiff tendons and no pennation angle. The tasks and physiological changes cover only a subset of possible ones, and baselines were run only on stationary tasks with fixed reset.
Figures
Abstract
Embodied agents in continuous control domains have had limited exposure to tasks allowing to explore musculoskeletal properties that enable agile and nimble behaviors in biological beings. The sophistication behind neuro-musculoskeletal control can pose new challenges for the motor learning community. At the same time, agents solving complex neural control problems allow impact in fields such as neuro-rehabilitation, as well as collaborative-robotics. Human biomechanics underlies complex multi-joint-multi-actuator musculoskeletal systems. The sensory-motor system relies on a range of sensory-contact rich and proprioceptive inputs that define and condition muscle actuation required to exhibit intelligent behaviors in the physical world. Current frameworks for musculoskeletal control do not support physiological sophistication of the musculoskeletal systems along with physical world interaction capabilities. In addition, they are neither embedded in complex and skillful motor tasks nor are computationally effective and scalable to study large-scale learning paradigms. Here, we present MyoSuite -- a suite of physiologically accurate biomechanical models of elbow, wrist, and hand, with physical contact capabilities, which allow learning of complex and skillful contact-rich real-world tasks. We provide diverse motor-control challenges: from simple postural control to skilled hand-object interactions such as turning a key, twirling a pen, rotating two balls in one hand, etc. By supporting physiological alterations in musculoskeletal geometry (tendon transfer), assistive devices (exoskeleton assistance), and muscle contraction dynamics (muscle fatigue, sarcopenia), we present real-life tasks with temporal changes, thereby exposing realistic non-stationary conditions in our tasks which most continuous control benchmarks lack.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.