Physical Intelligence's Chelsea Finn on the state of the art in robotics and scaling RL to the real world
At Y Combinator, Physical Intelligence co-founder Chelsea FinnDealroom has a profile for this one. Try Dealroom → laid out where general-purpose robotics stands two years into the company, whose mission is to let any robot do any task in the real world, and why real-world reliability, not cool demo videos, is the hard problem.
What the robots do now. Beyond last year’s laundry folding, Physical Intelligence’s models now wash a greasy pan, peel a carrot, make a grilled cheese, and slice a zucchini, generalising to rooms they have never seen before.
Physical AI is different. Unlike recommendation or chat systems where a human vets the output, robots act directly on the world, so they must run fully autonomously and make far fewer mistakes to be useful.
Reliability is the bar. Finn grounds it in making espresso: precise porta-filter insertion, spill-free cup handling and accurate timing, all at 90%+ reliability, so no one has to babysit the robot.
Toward self-improvement. Manual data-tuning plateaus; the goal is an RL-style system that automatically seeks out where it needs more data or supervision, iterating toward 99%+ reliability.
RL is costly in the real world. Language-model RL (PPO, GRPO) uses millions of attempts run on compute; one million one-minute robot trajectories would mean ~700 robot-days, so the recipe must be far more sample-efficient.
Human-in-the-loop recovery. Rather than waste robot time on dead-end trajectories (e.g. folding two boxes stuck together), a human teleoperates to show recovery, or the episode terminates early, keeping data efficient.
Foundation-model recipe. Progress came from combining large-scale real robot data via teleoperation with pre-trained vision-language models (e.g. PaliGemma); scale is “necessary but subordinate to solving the problem.”
Reason for optimism. Finn points to Waymo passing a quarter-million weekly autonomous rides as proof that trustworthy, fully autonomous ML systems can operate in the physical world.
Read more: Y Combinator, “Chelsea Finn: This is the State of the Art in Robotics” (YouTube).