Chelsea Finn: This is the State of the Art in Robotics

Video summary

Chelsea Finn says Physical Intelligence’s PIO7 controls diverse robots out of the box, while reinforcement learning doubled throughput.

At Startup School 2026, Physical Intelligence cofounder Chelsea Finn explained why useful robots must work autonomously and reliably, not just perform impressive demonstrations. Reinforcement learning with human interventions and a general value function helped robots learn from failures, double task throughput, and make espresso with over 90% success; one latte policy ran for 13 hours. Adding short-term video and longer text-based memory enabled a robot to clean a kitchen autonomously for 10–15 minutes. Finn then introduced PIO7, a single model trained on diverse robot demonstrations, robot experience, human videos, and web data. It matched or exceeded task-specific specialists out of the box and showed compositional generalization, including folding clothes on a robot platform with no folding training data. In the Q&A, she said robotics’ ChatGPT moment may be a few years away, emphasized real-world robot experience as essential training data, and described open-source models and hands-on projects as paths into robotics.

Chapters

1. 0:00The State of Physical Intelligence: Robots Wash Pans, Peel Carrots, and Make Sandwiches
2. 1:23What It Takes to Make Robots Useful: General-Purpose Models Must Act With Fewer Errors
3. 5:11The Reliability Problem: Espresso Robots Need Over 90% Success Without Babysitting
4. 7:43Reinforcement Learning for Robotics: 700 Robot Days for One Million Trials
5. 9:35Learning From Failures: Human Recovery and a Shared Value Function
6. 12:43Training Robots to Improve Themselves: A Latte Workflow With Human Collaboration
7. 14:21Can a Robot Work for 13 Hours Straight? RL Doubles Throughput Across Real Workflows
8. 17:36Why Robots Need Memory: Compressed Context Enables 10–15-Minute Kitchen Cleaning
9. 21:22Building a General-Purpose Robot: From 2012 ImageNet to One Out-of-the-Box Model
10. 25:02From Fine-Tuning to Out-of-the-Box Models: Generalizing Beyond Avocado-Chair Examples
11. 27:35Training on All the Data: PIO7 Folds Shirts, Drills Screws, and Replaces Trash Bags
12. 30:20One Model That Beats the Specialists: PIO7 Matches or Outperforms Fine-Tuned PIO6
13. 31:21Compositional Generalization: PIO7 Handles Air Fryers and Folds Shirts on an Unseen Robot
14. 37:49The GPT Era of Robotics: Ultra and Weave Deploy PI Models Beyond Demos
15. 39:49Q&A: ChatGPT-Like Robotics, PI0 and PIO5, and PhD Versus Industry
16. 45:09Q&A: Robot Experience, Teleoperation Data, and the Limits of Human Videos
17. 49:09Q&A: Open-Source Robotics, Joint-Position Control, Imagination, and RL Speed
18. 54:16Q&A: Pinwheel Hand Transfer, Vegetable Slicing, and Entering Robotics