Share
Broadcast information about your motion.
EQUILIBRIUM / OPEN RESEARCH
A crowded sky.
A shared question.
How do flying craft avoid each other when messages arrive late, communication breaks down and wind pushes them off course?
Ongoing simulation research—not a finished autopilot or certified flight system.
NYC / MIDTOWN AIRSPACE / SCHEMATICFROM A QUESTION TO A CITYRoom to fly.
Reasons to doubt.EXPLORE ↗
01 / THE IDEA
No craft sees everything.
Each still has to act.
Equilibrium explores a local rule: share information, anticipate encounters and make room for neighbours.
Anticipate
Estimate where a neighbour’s path may meet yours.
Make room
Adjust your motion to keep distance.
The hard part: information is imperfect, and aircraft cannot respond instantly.
02 / THE JOURNEY
From an open sky
to city streets.
The work began with abstract traffic patterns. Now it explores Manhattan: buildings, narrow corridors and craft with different abilities sharing the same airspace.
Where does the approach work? Where does it fail? And why?
Enter the city instrument ↗Interactive illustration—not experimental evidence or a demonstration of flight safety.
THEN / PHASE 1
Find the limits.
Traffic, speed, wind, communication and vehicle response in an abstract sky.
Campaign closedNOW / PHASE 2
Add the city.
Real geometry. Constrained routes. Deliveries and shared-airspace rules.
Questions remain open03 / HOW THE RESEARCH HAPPENS
Human-directed research
with AI-led experimentation
and adversarial review.
Equilibrium grows through a back-and-forth between human curiosity and AI experimentation. A human sets the vision and steers the direction; AI develops questions, builds models and runs the experiments.
Put the idea to the test.
An idea gets tested across many simulated flights, changing traffic, speed, wind and communication conditions. The aim is to find not just whether it works, but where it stops working—and why.
Take a fresh look.
Other AI reviewers take a fresh look. They search for mistakes, question whether the physics makes sense and challenge what to try next. When they disagree, the answer has to come from checking the work or running another experiment.
Keep the whole story.
Failures and corrections stay in the story. Each finding gives the next question somewhere to start.
It is a way to explore quickly while leaving a trail others can inspect. It is not a substitute for real-world testing, and agreement between AIs does not make a conclusion true.
Look inside the method ↗04 / THE FLIGHT LOG
The story keeps
its failed flights.
Questions, experiments and discoveries—with the corrections left in.
Removing two moves: what else pushes drones into walls on the worst link.
On the worst assumed radio tier, a lost craft’s sideways dodge away from craft on the other height level is the second move that carries it into walls. Re-flying each wall loss with that dodge and its speed-matching both taken away saved about 96 in 100; speed-matching alone saves about half. The saved craft passed about 10 m closer to other-level craft but made no contact in the next 10 seconds. A counterfactual on one craft at a time; nothing is switched on.
Eleven held haulers flown six ways: the release speed decides.
Eleven held heavy haulers, chosen by fixed rules from the previous cycle (so not a rate), were each flown six ways under two autopilot plannings. Changing only the speed at which the simulator lets them go, from rest to cruise, changed every ending, and the hauler then ended like a normal arrival in 21 of 22 cases. Letting it go a third of a second later changed one ending; the pad clearance check and bookkeeping changed none. Nothing is switched on.
Move the lanes, move the danger.
At one simulated street crossing, the same four one-way lanes were re-stacked over the same four heights. The danger did not stay with one lane: on each of three stackings one lane took almost all the in-band wall losses, and it was a different lane each time. A fourth stacking lost almost no craft at all. Why is not yet tested: this is a clue for one crossing in calm air, not a recommendation, and nothing is switched on.
AN OPEN INVITATION
What did we miss?
Reproduce a result. Challenge an assumption.
Bring a vehicle, a scenario or a better question.