Watch a flylearn to play.
Follow its feet. Watch the commands. See what happens in the world.
THE BODY MOSSLIGHT VALLEY
Awaiting connectionTry. Fail. Adjust. Try again.
AN ADAPTATION STORYHundreds of attempts, different paths. Early runs hesitate, backtrack and explore. Jump timing varies each attempt; exploration settles as experience grows.
First explore. Then use the outcome to adjust the next attempt.
Expect hundreds of attempts before the first success. Seeded exploration varies movement and jump timing. Timing changes after actual failures using an explicit error-correction rule. This is controller adaptation, not biological or neural learning. No successful outcome is prerecorded.
What body comes next?
THE NEXT EXPERIMENTToday, a leap. Next, a swim, a first step, or flight through an unfamiliar world.
A tiny body. A remarkable brain.
We are building a complete embodied system: a mapped brain, an articulated body and a loop of action, feedback and adaptation.
Mapped brain neurons
Reconstructed in the FlyWire adult female fruit fly brain atlas—the biological foundation for our full-brain integration work.
Mapped synapses
Connections linking the mapped neurons. This wiring diagram provides a structural basis for building and studying neural circuits.
Many possible bodies
Connect a shared neural system to different bodies and environments. First Leap develops the body, action and feedback loop; full connectome integration is in development.
Biological figures: Dorkenwald et al., Nature (2024) ↗ · Adult female Drosophila melanogaster brain, not the entire nervous system.
Built from anatomical geometry.
FLY—001 uses NeuroMechFly / FlyGym body meshes and a joint hierarchy: head, thorax, abdomen, six articulated legs and wings. Three.js renders the model in your browser, with added materials, eye textures and bristles.
Joint rotations place the front feet over the controls. The same action signal moves the game character and animates the fly. The movement makes each command visible as it happens.
Explore the open-source body model ↗Small rules. Changing timing.
The controller reads the character’s position, ground contact and distances to the step and gap. It selects one of four actions: idle, left, right or jump. In this build, those observations come directly from game state.
Each session has a random seed. Hesitation, backtracking, exploratory jumps and timing variation create different attempts. Exploration decreases as the attempt count rises.
Errors leave a small memory.
A failed attempt before clearing the step increases the step-jump trigger by 0.08 pixels. A failure after clearing it reduces the gap-jump trigger by 0.20 pixels. These small corrections carry into the next attempt.
This build uses engineered parameter adaptation to develop and test the learning loop. Success emerges from movement and collision physics, with a different exploration path in each session.
One attempt, from observation to feedback.
- 01 / ReadPosition, obstacles and contact
- 02 / ChooseTiming rules + seeded exploration
- 03 / ActGame input + synchronized 3D pose
- 04 / MeasureProgress, collision, fall or goal
- 05 / AdjustRetain timing for the next attempt
The game advances at 30 simulation steps per second. The 3D scene renders separately. Both run locally in this browser.
Building toward the complete system
Save position retains the game state, random seed, controller timing and attempt history in this browser. Download run exports the experiment record. These records make experiments reproducible as the system evolves.
The next integration stages connect the brain graph to neuron dynamics, sensory encoding, motor decoding and learning rules, with validation at each step. Our neural research prototype develops the computational side alongside this body-and-control experiment. The long-term aim is one continuous system that can perceive, act and adapt across different bodies.