NEURODRIVEBIOLOGY × MACHINE

BIOLOGICAL INTELLIGENCE / EXPERIMENT 003

NEURODRIVE

A tiny brain. Learning to drive.
An embodied learning experiment. Observe every input, error and breakthrough.

SUBJECT ONLINE001DROSOPHILA MELANOGASTER
Awaiting first attemptATTEMPT 001 / 500
02 / DRIVING ENVIRONMENTORCHARD RING · DRY
NEURODRIVE / LIVE TELEMETRY0.0%LAP PROGRESS
LEARNING STARTS AT ZERO.

No racing line supplied.
No victories scripted.

SPEED0 km/h
ATTEMPT TIME0.0 s
BEST PROGRESS
EXPLORATION94%

A fresh controller. A curved road. Let's see what experience changes.

FROM CRASHES TO CORNERS

Progress has a paper trail.

Every point is a real attempt in this simulation.
Some days, learning looks like going backwards.

LAP DISTANCE / ATTEMPT● EACH RUN ━ ROLLING 20

THE SCOREBOARD

Completed laps0
Off-track exits0
Fastest clean lap
States experienced0

20 held-out starts per controller. Same track and surface, learning frozen.

SESSION LOG / LATEST ATTEMPTS
ATTEMPTOUTCOMEDISTANCETIMEEXPLORATION
Start the engine to open the notebook.

ONE BODY. ANOTHER WAY TO LEARN.

The fly is the driver.
Experience is the instructor.

01 / OBSERVE

Read the road.

The controller observes lateral position, heading error, road curvature and speed. It receives no ideal racing line and no prerecorded steering sequence.

02 / ACT

Make a decision.

Nine combinations of steering and throttle move a simplified vehicle. The anatomical fly's wheel and pedals display those same inputs in real time.

03 / ADAPT

Remember the outcome.

Tabular Q-learning rewards forward progress and penalizes leaving the track. Action values persist between attempts; outcomes emerge from the simulation.

Protocol & biological context

This browser study uses an engineered Q-learning controller and road-relative vehicle physics, not a running whole-brain connectome. NeuroMechFly / FlyGym anatomy supplies the articulated fly. The FlyWire atlas of 139,255 neurons and 54.5 million synapses informs our longer-term biological integration work.

Fixed 0.12-second model steps. An attempt ends on a full lap, an off-track exit, or 216 simulated seconds. A fresh study resets memory. Human practice is isolated. Evaluation freezes learning and tests both policies with 2% exploration on held-out initial conditions. Faster playback changes wall-clock speed only.

Inspect the model ↗ · Previous foraging study ↗

COMMUNITY-FUNDED EXPERIMENTS
Planned for Robinhood Chain: a share of token trading fees supports future experiments, with holders helping choose research directions. Funding and governance are in development.