explainer
Physical AI from a data perspective
What changes when you move from tabular pipelines to perception, action and simulation — without pretending to be a robotics engineer.
Starting point
A data background already knows provenance, evaluation, drift, cost, latency. Physical AI asks the same questions of sensors, policies and simulators. It is not a magic second career. It is another kind of signal.
Useful vocabulary (not expertise)
- VLM — a model that ties image and language.
- VLA — a see → reason/speak → act loop.
- ROS2 — robot middleware, not practised here in public yet.
- World models / sim — where you would rather evaluate without breaking hardware.
What this is not
No professional robotics experience. No published ROS2 run. PhotoMind (personal photo intelligence) is still private: multimodal intent, not Lab evidence until it is opened.
Lab link
The First Physical AI notes lab is the entry. A first pixel run is public: A measurable vision baseline from a data seat (UCI digits 8×8). Decision Copilot and Knowledge Copilot remain data / agents / RAG evidence. Physical AI is the direction, not the current résumé.
Later, SignalLab
Separate research acceleration (VLA papers) from real adoption (repos, products, jobs). A tweet is not a deployment.