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

影片總結

Chelsea Finn shows Physical Intelligence robots achieving 90%+ espresso success, 13-hour autonomy, 2x RL throughput, and generalization across tasks.

At Startup School 2026, Physical Intelligence cofounder Chelsea Finn explains why useful robots need far more than impressive demos: they must operate autonomously, reliably, and for long periods. Reinforcement learning with human interventions and a general value function doubled robot throughput, while espresso-making exceeded 90% success and ran continuously for 13 hours. Memory at multiple timescales enabled a robot to clean a kitchen autonomously for 10–15 minutes. Finn then introduced the single-model PIO7 system, trained on diverse robot demonstrations, autonomous rollouts, human videos, and web data. PIO7 matched or exceeded fine-tuned specialists out of the box, generalized to unseen appliances such as air fryers and new robot platforms, and showed strong compositional learning. She argues robotics is entering its GPT era, but deployment, hardware, data, speed, and reliability remain major challenges.

章節

1. 0:00The State of Physical Intelligence: Physical Intelligence’s Laundry and Kitchen Robots
2. 1:23What It Takes to Make Robots Useful: From ChatGPT to Autonomous Physical AI
3. 5:11The Reliability Problem: Making Espresso Above 90% Reliability
4. 7:43Reinforcement Learning for Robotics: Reaching 99%+ Reliability Efficiently
5. 9:35Learning From Failures: Teleoperation and General Value Functions
6. 12:43Training Robots to Improve Themselves: Latte Preparation With Human Collaboration
7. 14:21Can a Robot Work for 13 Hours Straight? Latte, Boxes, and Clothing Tasks
8. 17:36Why Robots Need Memory: 10–15-Minute Autonomous Kitchen Cleaning
9. 21:22Building a General-Purpose Robot: From ImageNet to Compositional Generalization
10. 25:02From Fine-Tuning to Out-of-the-Box Models: Avocado-Chair Compositionality
11. 27:35Training on All the Data: The PIO7 Single Robot Model
12. 30:20One Model That Beats the Specialists: PIO7 Matches Fine-Tuned PIO6
13. 31:21Compositional Generalization: Air Fryers, New Robots, and Diverse Data
14. 37:49The GPT Era of Robotics: Physical Intelligence Reaches Real-World Deployment
15. 39:49Q&A: Robotics’ ChatGPT Moment, Generalist Policies, and PhDs
16. 45:09Q&A: Robotics Data Requires Real Robot Experience and Hardware
17. 49:09Q&A: Open Source Models, Joint Control, Imagination, and Speed
18. 54:16Q&A: PIO7 Transfers Pin-Wheel Skills and Jenny Enters Robotics