parametric spaceBuild the experiment
Place food, barriers, and hazards. Set movement, rewards, episode length, sensory windows, and fixed neural readouts.
RESEARCH / NEURAL SIMULATIONPS — EXPERIMENT 02
THE STONKFLY FORKA small fly.
A small fly.
A whole world
to explore.
A real connectome inside an experimental world. Build a scenario, shape its inputs, and inspect what happens between sensation and action.
Open the simulationInteractive browser demo · real training requires compute
DROSOPHILA / BODY INSPECTOR3D ASSET
FlyBody anatomy · articulated in our labDRAG TO ROTATE
166,700neurons in the retained graph
25.6Mretained connections
5task modes in the training lab
1shared experiment interface
01 / WHAT WE ADDEDFrom a single task
From a single task
to an experimental workspace.
Our fork of Stonkfly extends the original trading experiment into configurable environments and custom behaviors.
Inspect the whole loop
Watch the 3D environment beside the fly’s sensory image, neural activity, selected actions, and body constraints.
Practice, save, compare
The Python lab supports continuing checkpoints, frozen-memory runs, and matched tests before and after practice.
02 / INSIDE THE SIMULATION
Sensation becomes action.
01 / OBSERVE
Visual scene
Environment geometry becomes a 160 × 90 sensory image.
02 / PROPAGATE
Retained connectome
The full graph runs in the Python neural engine.
03 / DECODE
Fixed readouts
Selected neuron groups propose named actions.
04 / ACT
World & feedback
Constraints limit execution; task rewards feed reinforcement.
Movement and sensory projection are engineered interfaces. The body animation is a bounded kinematic illustration, not calibrated biomechanics. Synaptic changes alone do not establish successful learning.
03 / FROM BROWSER TO COMPUTEThe same workspace.
The same workspace.
An honest view of what runs.
The website reuses the fork’s actual lab interface and fly assets. Its runtime adapter keeps the existing experiment configuration and API contract.
Explore the controlsAVAILABLE NOW
Browser environment
Design and export scenarios, inspect the fly, and drive the arena manually. Movement, swept collisions, food collection, and reward terms follow the Python arena rules.
COMPUTE CONNECTION
Full neural training
A server connection is prepared for the real engine. Running the connectome, learning, checkpoints, assays, and live website tasks require a provisioned compute gateway.
NEXT
User training sessions
Isolated jobs, compute limits, durable run storage, and user access are the next integration step.