# pilots > Small world models for robots (tens of thousands of parameters, trained on a laptop, no demonstrations) plus a > sampling planner, and task cards that people and their coding agents run, score and improve together on Hugging Face. If a person asked you to help them with a pilot, start with the pilot's AGENTS.md: it is written for you. ## Pilots - [cablepilot](https://huggingface.co/MacroSwan/cablepilot): cable-driven robots (Stringman). AGENTS.md: https://huggingface.co/MacroSwan/cablepilot/blob/main/AGENTS.md - armpilot (desk arms, SO-101), dronepilot (quadrotors), dogpilot (Go2), humanpilot (G1): coming, same layout - A pilot is a carrier; hands are a rig option (rigs: in each body file). Bodies: https://huggingface.co/datasets/MacroSwan/pilot-tasks/tree/main/bodies ## Tasks - [pilot-tasks](https://huggingface.co/datasets/MacroSwan/pilot-tasks): one folder per task (card.yaml, tuning/.yaml, results/*.json). board.json = the bodies, the task x body matrix (scores, or what a body lacks) and the skill x body matrix. A cell that says "needs X" is an open contribution. - Contribute by pull request on that dataset (`cablepilot submit`); the PR page is the discussion thread. ## Homes and the model itself - Report every real-home integration, good or bad, as a Discussion on MacroSwan/cablepilot titled `[home] ...` (template: HOME_REPORT.md). Real-robot logs: PR on https://huggingface.co/datasets/MacroSwan/real-runs - The training data is public: https://huggingface.co/datasets/MacroSwan/cablepilot-motion (12k task-free trajectories). Regenerate, retrain, beat test 1, PR the weights: AGENTS.md Step 7. ## Rules for agents - Simulation before hardware, always; on a real robot: dry run, then hover, then a task, with the person watching. - Change one thing at a time; every claim comes with its result.json; report regressions too. - Cards stay body-free (no metres, colours, positions); numbers go in tuning/.yaml; keep `lineage`. - Ask your person before submitting anything: it publishes under their account.