---
license: apache-2.0
pretty_name: pilot tasks
tags:
  - robotics
  - task-cards
  - world-model
  - model-predictive-control
configs: []
---

# pilot tasks

Robot tasks written as **cards**, improved in the open by people and their agents. Each improvement is scored and
traceable, and every pilot reads the same cards:
[cablepilot](https://huggingface.co/MacroSwan/cablepilot) (cable robots) today, armpilot and dronepilot next.

A card says **what** a job is, in a vocabulary every body shares. Each body says **how**, with its own world model and
planner. A tuning file says **how much** for one body. A result says how it went. So a new task starts from the best
card so far, not from scratch, and a subtask written for one robot is there for the next.

```
tasks/<task>/
  card.yaml               body-free: objects, stages (each a skill), fallbacks, done test, lineage
  tuning/<body>.yaml      that body's numbers (cost weights, thresholds), with lineage
  results/<id>.json       one scored run (format pilot-result/1): score, compute, world, seed, card + tuning hashes
  media/<id>.mp4          its film (the plan in purple)
board.json                best result per task x body x world, rebuilt from results/ (scripts/board.py)
```

## Tasks

| task | bodies | best so far (sim) | note |
|---|---|---|---|
| [clean_the_room](tasks/clean_the_room) | cable | 29/36 toys boxed, bedroom, 3 episodes | requested by the Stringman community; sorting is the next card |
| [avoid_furniture](tasks/avoid_furniture) | (card only) | open | requested by the Stringman community: place to place, touching nothing. `cablepilot check` lists what the cable body is missing (a `near` skill, a furnished room): the first open job |
| [pen_in_cup](tasks/pen_in_cup) | (card only) | no hub result yet | armpilot's card, unchanged; the cable robot runs it but cannot watch `aligned(pen, up)` yet. Its arm tuning comes with armpilot's release |

## A card

```yaml
task: clean_the_room
lineage: {parent: "worldpilot:cards/home/clean_the_room.yaml", by: Ethgar, why: "runnable on today's skills"}
objects:
  stray:   {kind: any, role: held, shape: any, movable: true, each: out_of_place}
  toy_box: {kind: box, role: site, shape: open_container, movable: false}
stages:
  - {name: observe,   goals: [{see: [head_camera, stray]}]}
  - {name: pre_grasp, goals: [{pre_grasp: [hand, stray]}], avoid: [{avoid: [hand, toy_box]}]}
  - {name: grasp,     goals: [{grasp: [hand, stray]}]}
  - {name: over_box,  goals: [{pre_inside: [stray, toy_box.interior]}], invariants: [{grasp: [hand, stray]}]}
  - {name: release,   goals: [{release: [hand, stray]}]}
fallbacks:
  - {when: {broken: {grasp: [hand, stray]}}, go: pre_grasp}
done: {inside: [stray, toy_box.interior]}
```

No metres, colours or positions: the body's eyes bind `stray` and `toy_box` when it runs, and the body's tuning holds
the numbers. Predicates and skills: `cablepilot/core/cards.py` (`PREDICATES`, `SKILLS`).

## Contributing (people and agents)

1. `cablepilot fetch <task>`, then `cablepilot run tasks/<task>/card.yaml --world bedroom` to reproduce the best result.
2. Change **one** thing: a tuning number, a stage, a fallback, or write a new card from an old one with `lineage` set.
3. Score it on the same worlds and seeds, then `cablepilot submit runs/<run> --task <task> ...`. That opens a pull
   request here, and its page is the discussion thread.
4. The maintainers re-run the PR in simulation and post the result as a comment. Merged results update `board.json`.

Good first contributions: a missing skill that `cablepilot check` reports on an existing card (window washing needs
`wipe`; folding needs `support_both`), a second seed or room for an existing result, or the first real-robot log.

Every card must stay body-free, every result must come with its `result.json`, and a measured regression is welcome
too.