• updated 2026-09-11 · 9 min read The data, the rig, and the exam The reference piece behind the runs: the rig, the two corpora, the exam locked before any model trained, and what the buyers themselves say they need from footage like this.
• updated 2026-09-11 · 4 min read What we learned training GR00T N1.7 One run on NVIDIA's model, the only family pretrained on first-person human video, which comes closest to the baseline of anything we have trained.
• updated 2026-09-11 · 4 min read What we learned training SmolVLA Three runs on the compact model, which reads our footage better than anything else we have trained and holds the programme's strongest scores.
• updated 2026-09-11 · 3 min read What we learned training π0.5 Four runs on the reference model: the footage trains it cleanly, label quality outranks label quantity, and volume was never the bottleneck.
• updated 2026-09-11 · 2 min read What each generation bought The pipeline has shipped twice, every measure of the data moved up between generations, and the scores moved with it.