Fly connectome with 138k neurons drives a robot duck's escape reflexes in sim
willknight · x · 2026-11-06
Someone connected the complete FlyWire connectome of an adult fruit fly (Shiu et al. 2024 — 138,639 neurons, 54M synapses) to a Microduck robot in a MuJoCo twin, letting the fly's descending neurons drive it — with zero training or tuning of the brain:
- ToF sensor → the fly's looming detectors (LPLC2/LC4): a ball thrown at the duck's face fires the giant fibre escape neuron on that side, and the duck flinches via existing robot.sound/robot.look/robot.pose/robot.move intents; a hand moving at petting speed (0.2 m/s) does not trigger it.
- Camera → the fly's motion detectors (T4/T5), each at its own visual column. In a rotating striped drum, head neuron DNp15 turns the duck's head with the stripes at a closed-loop gain of 0.3–0.4, and the walking body follows.
The project also fixes a duck-sim bug where the duck couldn't actually walk (pre-v5 policies loaded on a model without backlash). Code is on GitHub (duckophila); the README clearly separates what's the fly's from what's glue code.
More from Embodied
- Lamport Robotics' Pincer1: a Six-Legged Robot Wallet That Signs Only When You're in the Room — R4nsack · 2026-11-06
- UCLA's FlowPilot uses human preferences to cut sidewalk robot interventions by 40% — zhoubolei · 2026-11-06
- RoboColosseum tests 5 VLA policies across 4 embodiments; environment diversity is the new bottleneck — DJiafei · 2026-11-06
- RoboColosseum: 20,000+ runs, 100+ robots, rankings flip by task — DJiafei · 2026-11-06
- Using Dance Choreography Tool Swing Choreo to Design Robot Motions — andrew_n_carr · 2026-11-06
- Cartwheel lands on founder-generated Emergent startup list for kinematic world models — andrew_n_carr · 2026-11-06