A fruit fly simulation detailed enough for Nature

What it does

flybody is a full-body physics model of the fruit fly Drosophila melanogaster for the MuJoCo simulator. It packages the fly as a reinforcement-learning environment with tasks for walking imitation, flight, and vision-guided flight, supported by optional distributed training via Ray and DMPO agents. The model is anatomically detailed and intended as a platform for fruit fly biophysics and embodied sensorimotor research.

The interesting bit

The model treats insect locomotion as a high-dimensional control problem, with a 59-dimensional action space for walking alone. It is built with enough anatomical fidelity to appear in Nature and is meant to move fruit fly neuroscience from observation into embodied simulation.

Key highlights

• Anatomically detailed Drosophila model for MuJoCo physics
• RL environments for walking, flight, and vision-guided flight
• Distributed training support via Ray and DMPO agents
• Published in Nature (2025) by Google DeepMind and HHMI Janelia
• Tiered installation: core physics, ML inference extension, or full distributed training extension

Caveats

• Core installation excludes TensorFlow and Acme, so policy rollouts and training require opting into the ML or Ray extensions
• Rendering and GPU library paths may need manual environment configuration
• Scope is tightly focused on fruit fly biophysics; it is likely overkill for generic locomotion problems

Verdict

A strong choice for biomechanics and neuroscience researchers who need an embodied, physics-based fruit fly. If your RL research does not involve insects or biophysics, this is likely not your benchmark.