lab
First Physical AI notes from a data seat
What should a data profile learn first to approach Physical AI?
Hypothesis: Data questions (provenance, eval, cost, drift) transfer. Vocabulary (VLM, VLA, world models) comes before ROS2.
Setup
No robot, no published sim. Reading plus a mental map. PhotoMind (library, clustering, search) stays a private personal project.
Result
A useful map:
- VLM — see and describe.
- VLA — see → language → action.
- ROS2 / sim / world models — not started in public.
The same questions as on a CSV apply: where the signal comes from, how error is measured, what it costs.
First public pixel run: lab A measurable vision baseline from a data seat (UCI digits 8×8, majority vs centroid).
Failure / limit
No perception-action run. ROS2, simulation and world models are not started here. The digits baseline is not a VLM.
Next
Change dataset or task before stacking robotics vocabulary. Tie to the explainer Physical AI from a data perspective.