What it is: models that understand movement in a body — and the thread with his name on it.
Why it is in this archive
Gate: it is one of his open questions, and questions are entities from birth. The standing
question is What sensors does a model need in order to understand the physical world?, and register idea 266-01 carries the NAME ON IT tag.
The gap, stated precisely
The robots learned martial arts before humans wrote it down. KungfuBot (arXiv 2506.12851) trains on real hardware, while the human motion data is acknowledged-missing in the literature.
⚠️ And the reframe that matters — the original claim "nobody wrote martial arts down" is false and trivially refuted. The correct spine: everybody wrote it down; nobody wrote it down the same way. Same named kick, three schools — one thrusts, one pulls, one snaps, all three correct. A taxonomy that cannot represent legitimate variation is one lineage asserting itself.
What he already has that nobody else does
- YOLO pose on the studio M2 — Ultralytics pose estimation
- ~12h Burton Richardson seminar video + the instructor course filmed end to end — expert coaching language paired to technique. Every public martial-arts dataset is motion-only. ⚠️ Burton taught it; Gonçalo hosted. Rights are a relationship call, not an admin task.
Where it came up
- Moonshots #266 — the sensors question, and the taxonomy-of-fighting question
- Moonshots #284
24:25— "a digital twin of your ability to move"; biomechanically correct video
⚠️ The honest caveat
Generated video is atmosphere, not technique — a generated dojo already rendered a named throw wrong. The defensible path is pose estimation on his own footage, not a model inventing movement.