Robotics: why now?
summary
Physical Intelligence demonstrates a pipeline of teleoperated data collection with leader-arm teleoperation and cloud annotation that yields 10,000 hours of successful episodes (up from 3,800), trains vision–language–action models with an action-expert transformer and diffusion for high-frequency controls, and releases PI zero as an open-source dextrous policy. PIO5 shows open-world generalization, running continuous autonomous episodes up to ten minutes in unseen homes and even controlling a distant robot to make coffee end-to-end.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 0:40Robotics before
- 1:24Robotics today
- 2:28What are VLMs
- 3:26Engineering challenges
- 5:39How does it work
- 7:39How far have we gotten
- 9:22Industry trends
- 12:43Long horizon tasks
- 14:22Subdividing tasks
- 15:37Why is this important
related talks
16mReal-World Robotics as the Next Frontier for AI?
01h 35mDeep Robotic Learning
29mBridging AI & Robotics: Exploring the LeRobot Library
14mTowards Unobtrusive Physical AI: Augmenting Everyday Objects with Intelligence and Robotic Moveme...
19mThe Physical Turing Test: Solving General Purpose Robotics
15mLeRobot – Lowering the entry barrier to AI for robotics
17mBeyond Automation: AI and the Evolution of Robotic Intelligence
32mBeyond the Screen: How Embodied Manifestations of AI can Amplify People in the Physical World