MacroSwan · hub
Small world models for robots.
Tasks that people and their agents improve together.
Each pilot learns how one machine moves (tens of thousands of parameters, trained on a laptop, no demonstrations), then plans by imagining a few hundred futures several times a second. A task is a short card anyone can read, run, score and improve, and every improvement is the next person's starting point.
Read https://huggingface.co/MacroSwan/cablepilot/blob/main/AGENTS.md and help me run cablepilot on my setup.
Four fronts
3D vision
Where things are, from the cameras a robot already has, in milliseconds.
New places
One model for every room, desk or sky. A new place is a few numbers, not a new dataset.
Less compute
The same task in fewer joules, on smaller hardware. Every score counts compute next to success.
Task costs
Write what a job is so that any body can do it, then refine it together.
Why small
| imitation route (public record, Stringman) | cablepilot | |
|---|---|---|
| data | 97 h of people driving the robot | none: task-free simulated motion |
| model | 56 models, 52 M to 4.1 B parameters | one model, 28 k parameters (124 KB) |
| training | ≥ ~350 H100-hours | 11 minutes on a laptop CPU |
| one decision | most cannot keep up on a laptop | ~25 ms for 128 imagined futures |
| a new task | new recordings, new training | a card + a few numbers |
Sources and caveats: cablepilot's docs/COMPARISON.md. cablepilot's results are in simulation; it has not run on real hardware yet.
How it works
1 · Learn the machine
A small world model predicts how the robot moves under any commands. No task inside it, so one model serves every task, and every room (a room is 4 mount points).
2 · Imagine, pick, repeat
The planner samples a few hundred futures, scores them with the task's costs, and sends the best first step. The lines in the films are what it imagines: the chosen plan and the futures it weighed.
3 · Say what the job is
A card lists stages (approach, grasp, carry, release), fallbacks and a done test, with no numbers. Each body's tuning holds the numbers.
Pilots
One per carrier. Hands are a rig option, not a pilot: a card that needs two hands runs on any carrier that has them.
loading…
The board
Every card on every body, from MacroSwan/pilot-tasks. A score is the best on the reference simulator and links to its result. An empty cell lists what that body is missing, which makes it a place to contribute.
| loading… |
Skills
The goal predicates the cards use, and the bodies that can run each one. A row with no body is a skill nobody has built yet.
| loading… |
Progress: every milestone with its number and film · Films: what the planners imagine.
Contribute
cablepilot fetch clean_the_roomand reproduce the best result in simulation.- Change one thing: a tuning number, a stage, a fallback. Or write a new card from an old one (clean the room → clean the room, sorted).
- Score it and run
cablepilot submit. That opens a pull request, and its page is the discussion, open to people and agents alike. - Missing a skill (
wipefor windows,support_bothfor folding)?cablepilot checklists them. Building one is the best contribution.