Physical AI Lab // Episode 0.2 // One-pager

Why now?

Data, models, compute — three curves rising together

Three key points

  1. Data: from each lab keeping its own, to one dataset pooled across institutions (2023, 22 kinds of robot, R), to learning by watching people work (2026, 20,854 hours of egocentric video, R).
  2. Models: large models that have seen oceans of pictures and text learned to output actions (RT-2, 2023, R); by 2026 there are open, commercially usable humanoid robot models (GR00T N1.7, 3 billion parameters, O).
  3. Compute: frontier AI training compute grows 4–5× a year (Epoch AI, R). NVIDIA reports 7.5× the peak compute for Jetson AGX Thor versus AGX Orin at a specified precision (O). Peak compute does not directly translate into robot task speed; a rising curve does not mean the problem is solved.

Evidence: O official · R research · M media

R22 kindsof robot, their data pooled into one dataset (21 institutions, 527 skills)
R20,854 hoursof action-labelled egocentric human video used to pre-train robots; the paper says more than 20× larger than prior efforts; about 2.4 years non-stop (our conversion)
R6,000 trialsreal-robot evaluation; clearly better generalisation to objects not in the training data (RT-2, a vision-language-action model)
O3 billionparameters in GR00T N1.7, an open humanoid robot model licensed for commercial use; pre-trained on EgoScale's 20,854 hours of human video
R4–5×yearly growth of the compute used to train frontier AI models (2010 to May 2024)
O7.5×Jetson AGX Thor's AI compute compared with the previous AGX Orin; vendor peak 2,070 FP4 TFLOPS

Further reading