Finding food, over time.
SUCCESS RATE · LAST 20 TRIALSLearningNo memoryTrial number →
Each point uses up to 20 completed trials. No example data are shown.
BEHAVIORAL SYSTEMS / STUDY 002
A little hunger.
A lot to learn.
A digital fruit fly. An unfamiliar maze. One food source.
Watch experience turn wandering into a way forward.
THE FORAGING STUDY02 / ∞
SUBJECTFLY—001
METHODPaired learning trial
STATUS Browser simulation
Ready to exploreTRIAL 001 / 300
THE ENVIRONMENT
Find the food.
Learning subjectNo-memory controlFoodSEED 3
STEPS THIS TRIAL0 / 360
CELLS EXPLORED1
EXPLORATION RATE94%
LAST REWARD—
Both subjects begin at the same point. Only the learning subject retains experience.
Your manual exploration stays outside the study.
OBSERVATION → EVIDENCE
Does experience make a difference?
The same maze. Different memory.
LAST 20 TRIALSLearningControl
Food found——
Mean steps / trial——
Q-values retainedYesNo
Start the study to measure the difference. One maze describes this experiment, not all fly behavior.
A separate evaluation uses 5% exploration for both subjects, without further learning.
LAB NOTEBOOK
Every attempt leaves a record.
| Trial | Learning outcome | Steps | Control outcome | Steps | Exploration |
|---|---|---|---|---|---|
| No measurements yet. Your first run starts a fresh notebook. | |||||
BEHIND THE EXPERIMENT
Biology inspires the question.
Evidence shapes the answer.
We are building toward an embodied digital fruit-fly brain. This study tests one piece of that system: the loop between action, reward and memory.
FLYWIRE · ADULT FEMALE BRAIN ATLAS139,255
mapped neurons
THE CONNECTIONS BETWEEN THEM54.5 million
mapped synapses
Biological reference for our integration work. Dorkenwald et al., Nature (2024) ↗
A body you can observe.
NeuroMechFly / FlyGym anatomy supplies the head, wings and six articulated legs. The forelegs reach for a four-key control deck. Key presses use the actual maze commands; the keys release when the subject stops or finds food.
Body model source ↗Learning you can inspect.
A tabular Q-learning controller stores four action values per cell. Each step updates one value from the received reward and estimated future value. No route is supplied to it.
Q ← Q + α [r + γ max Q′ − Q]A result you can reproduce.
Both subjects share the maze, start, food, action budget and exploration schedule. Only one updates its action values. Exported records can be rerun from the seed and checked for an exact match.
Read the experiment model ↗Read the full protocol and current system boundaries +
What actually runs
The browser runs a discrete maze and Q-learning, with exact grid position as the observation. It does not run the full FlyWire connectome. Anatomical motion is procedural; the displayed action values and study results come from the controller.
Rewards and timing
Moving costs 0.025 reward; wall contact adds a 0.2 penalty. Food provides +12 and ends the trial. Moving closer within the four-cell local food-scent region adds +0.04. The full maze distance field is not an input to the policy. Each subject has at most 360 actions.
Controls and reproducibility
Learning rate α = 0.18; discount γ = 0.97. Exploration follows 0.06 + 0.88 / [1 + ((trial−1)/65)²], identically in both groups. Seeds control maze generation and action sampling. Playback speed changes how fast steps are shown, not the results.
Interpreting the evidence
Trials share a maze and are not independent biological replicates. Mean steps include unsuccessful 360-step trials. Success means reaching food, even if it happens by chance early on. Repeated success and shorter paths are stronger signs of learning. Try additional seeds before generalizing.
WHERE INSTINCT BECOMES EXPERIENCE.
What should we ask next?
Changing the environment changes the question.
Help shape the experiments that follow.
Choose a direction. This is a local preference, not an on-chain vote.
COMMUNITY-FUNDED EXPERIMENTS
Planned for Robinhood Chain: a share of token trading fees supports future experiments, with holders helping choose research directions. Token funding and governance are in development.