Brain
Open base models, post-trained on our own deployment data, with a stated reliability target.
Where we start
We do not train a base vision-language-action model from scratch. NVIDIA's GR00T is released under a commercial licence and Physical Intelligence's π models are open. The base layer is becoming a commodity, and the leverage is in what happens after it.
The last mile
Task-specific post-training on data from our customers' sites is what crosses the economic threshold. Physical Intelligence's reinforcement from corrections, Generalist's one hour of task data and DeepMind's fewer than 200 demonstrations all point the same way. Our target for any task we sell is 95 percent success without intervention, and we publish the number.
Compute
Training runs on our own GPUs in Singapore. We keep a clean capitalisation table so that access to training hardware is never in question, and we state that plainly to investors.
What we publish
Evaluation results, including failures, in the Brain category of Insights.