Overview

A founder from Physical Intelligence walked through their push toward general-purpose household robots that can follow natural commands and handle long, dexterous tasks—think folding laundry, boxing, wiping spills, lighting a match, or “clean the bedroom”—in homes they’ve never seen. The recipe: tons of teleop data across many robots and environments, vision-language backbones that “speak robot” via continuous action heads, and broad pretraining that doubles performance and enables zero/low-shot transfer; they’ve open-sourced models that others are already fine-tuning. Near term, they’re focused on speed, safety, and robustness (with RL likely), and long term they see robot-generated real-world data becoming the next fuel for foundation models across many embodiments, not just humanoids.