“More complex and dexterous tasks may require the model to be fine-tuned to specialize it to downstream challenges.”
π0 starts from a 3 billion parameter VLM and drives robots at up to 50 commands a second, trained on data from 8 different robots. The laundry folding demo is the fine-tuned version, not the generalist. Physical Intelligence scores its own five tasks and compares against OpenVLA and Octo, which scored 0 and 0.043 on the easiest one. Beating a zero is a low bar to clear for “generalist.”