Flockinator - a microscope for emergent behaviour

Sixteen embodied agents, staged in sound, light, and language on a wall of nodes, so humans can spot what emerges between agents - then prove or refute the spot against sealed evidence.

Why this is useful: we experiment with intelligent agents to identify emergent behaviour — what a group does that no single agent was programmed to do. The point is practical: edge devices such as search-and-rescue quadcopters must act autonomously in the real world, where communications fail and sensors lie. A swarm that has learned when to trust, when to corroborate, and when to abstain is safer to deploy than one that has only ever flown in friendly conditions.

Interactive simulation only. Nothing here controls real hardware.

Start here: the full tutorial

New to Flockinator? Build a panel node with the Waveshare parts list, learn how swarms think between nodes, read the experimental method, browse scenarios from rescue quadcopters to fruit-fly neurons making music, or skip the hardware entirely and run the virtual wall.

Illustration of a small drone swarm searching earthquake rubble, one drone spotlighting a survivor
Illustration (not evidence): the search-and-rescue scenario the panel swarm simulates. Generated with Meta muse-image-1.0, 2026-11-06; real wall photography is below.

Flagship benchmark: messages help, corruption hurts more than silence

Seeded host experiment (field-signals-v1): teams of agents identify sectors from private cues plus peer messages. Team success = at least 3 of 4 agents correct.

Loading pooled result...

Muted/private baseline receives no peer messages by design (different information budget); tasks within a seed are not independent replicates. Reproduce from source: uv run python tools/reproduce/run.py --verify in the public repository.

The wall: watch trust fail and recover

Replay of one scored simulation (composite-defence world, seed 200): forged reports and a sensor modality that starts lying at contact 20. Two policies decide each contact. This is a visualisation of a host-scored run - the microscope view - not live nodes.

Both policies exclude reports labelled forged in this simulation; “verified” is label segregation, not cryptographic proof. E requires at least 2 survivor reports and a survivor-minus-hazard margin of 2. COMP learns modality weights and uses weighted thresholds of 1.5 and 1.0 respectively. This replay changes both weights and thresholds, so it does not isolate the effect of learning.

Animated montage of the sixteen real panel displays in flight-simulation mode
The actual panel animations the physical nodes play (flight-simulation display set, reduced for the web). Each replay cell below uses the same audited animation segments: scanning, sharing a report, checking, or doubting a claim.

Published results

Scripted demonstration (not a live run)

A 12-second scripted replay of a scored simulation, plus the real wall photographed at idle. See the existing 90-second physical wall footage for the display in motion. These are separate demonstrations, not a continuous physical attack, model review and intervention sequence.

The real 16-panel wall photographed at idle, every panel showing READY
The real wall at idle (READY / REMOTE START on every panel), captured on the lab camera rig, 2026-11-06. Panels animate only when a mission display is driven.