Training robots for a world they’ve never seen
About this episode
Chelsea Finn discusses how robots can learn to adapt to new environments and objects, emphasizing the importance of generalization in AI.
Useful robots won’t be programmed one task at a time; they’ll need to adapt to unfamiliar objects, environments, and robot bodies. Physical Intelligence cofounder Chelsea Finn (named last month to the TIME100 AI list for 2026) joins Reid Hoffman and Aria Finger to explain how general-purpose robot models learn to act in the messy physical world. She explores why lower-quality training data can strengthen a model, what months of failed laundry-folding attempts revealed, and how the same system can work across different robot platforms. A robot mistaking an oven for a drawer becomes a lesson in the strange line between useful generalization and obvious error. They also examine why robot demos can mislead, how hardware reliability can stall progress, and what it will take to move AI off the screen and into the world. Congratulations to Chelsea, recently named one of TIME’s 100 most influential people in AI: https://time.com/collection/time100-ai/2026/chelsea-finn/
People in this episode
Host: Reid Hoffman
Guest: Chelsea Finn
Topics covered
- robot training
- AI adaptation
- physical intelligence
- general-purpose robots
- machine learning
- robot platforms
Keywords
- robots
- AI
- machine learning
- generalization
- training data
- physical intelligence
- automation
Mentioned in this episode
Organizations: TIME
More episodes of Possible
- Can we defend against AI hacks? | Dozie Anazia
- Inside an AI-powered personal dashboard | Matthew Tiemann
- The whole studio is one guy | Jonathan Brazeau
- Making taxes fun with Pikachu and AI | Cadi Zhang
- How Sougwen Chung teaches robots to pause
- The artist using AI to sell brands | Joe Salvatore
Explore listener stats, chart rankings, contacts and more on the Possible podcast page.