Salk Institute for Biological Studies

Research

Computational models of biological motor control, neurorobotics, and dynamical systems analysis.

Featured

luc3d — 3D Multi-view Pose Annotation GUI Software · In Development

luc3d — Label Unification and Correspondence in 3D

Leonardis, E. J., Oline, S., Park, J., Maree, L., Afshar, S., Faulkner, A., Pereira, T. D.

A web-based multi-view 3D pose annotation GUI built with vanilla JavaScript and Three.js. Features multi-camera video playback, DLT triangulation, reprojection error visualization, cross-view tracking, and SLEAP-compatible import/export. Although browser-based, it loads locally mounted files quickly for a fluid user experience. No build system — pure static files.

Mouse Forelimb Imitation Learning
NeurIPS Workshop 2025

Massively Parallel Imitation Learning of Mouse Forelimb Musculoskeletal Reaching Dynamics

Leonardis, E. J., Nagamori, A., Thanawalla, A., Yang, Y., Park, J., Saunders, H., Azim, E., Pereira, T. D.

Our paper has been accepted to the NeurIPS 2025 — Data on the Brain & Mind: Concrete Applications of AI to Neuroscience and Cognitive Science Workshop. We utilize MIMIC-MJX which uses JAX and MuJoCo-MJX to elicit speeds of more than 1 million steps per second through the physics and RL environment. This enables massively parallel imitation learning of mouse forelimb musculoskeletal reaching dynamics, bridging high-fidelity biomechanical simulation with scalable reinforcement learning to study the neural computations underlying motor control.

arXiv 2025 · NeurIPS Workshop, San Diego, CA
MIMIC-MJX Framework
Preprint · Under Review at Nature Methods

MIMIC-MJX: Neuromechanical Emulation of Animal Behavior

Zhang, C. Y., Yang, Y., Sirbu, A., Abe, E. T. T., Warnberg, E., Leonardis, E. J., Aldarondo, D. E., Lee, A., Prasad, A., Foat, J., Bian, K., Park, J., Bhatt, R., Saunders, H., Nagamori, A., Thanawalla, A. R., Huang, K. W., Plum, F., Beck, H., Flavell, S. W., Labonte, D., Richards, B. A., Brunton, B. W., Azim, E., Ölveczky, B. P., Pereira, T. D.

A new collaborative preprint on our neuromechanical modeling framework, integrating stac-mjx and track-mjx for data-driven control of biomechanical bodies in physics simulation. MIMIC-MJX enables high-speed imitation learning from motion capture data through fully differentiable, GPU-accelerated simulation, allowing researchers to fit biomechanical models to real animal movement data at unprecedented scale.

arXiv 2025

Publications

Video

PiRat: rat-robot interaction project in collaboration with Janet Wiles' CIS Lab at UQ.

Affiliations

Salk Institute for Biological Studies