Chelsea Finn: Building Robots That Can Do Anything

The podcast features a presentation about developing general-purpose robots with physical intelligence, focusing on the challenges of creating robots that can perform various tasks in different environments. The speaker discusses the importance of scale and diversity in data collection, highlighting the pre-training and post-training recipe used to train robots for tasks like laundry folding, table cleaning, and coffee making. The presentation also covers how robots can succeed in new environments and respond to open-ended prompts by leveraging synthetic data. The speaker then answers audience questions about high-quality action data, the role of reinforcement learning, funding for robotics, the interplay of VLAs and world modeling, infrastructure layers, model sizes, and hardware and software evolution.

Outlines

Sign in to continue reading, translating and more.

Continue

Y Combinator Startup Podcast

The Challenge of Developing General Purpose Robots

Training a Foundation Model for Dexterous Tasks: Laundry Folding

Improving Laundry Folding with Vision Language Models and Transfer Learning

Generalizing to Unseen Environments with Diverse Data

Responding to Open-Ended Prompts and Interjections with Synthetic Data

Q&A: High-Quality Data, Funding, Infrastructure, and Model Size

Chelsea Finn: Building Robots That Can Do Anything

Y Combinator Startup Podcast

00:00The Challenge of Developing General Purpose Robots

The Challenge of Developing General Purpose Robots

04:15Training a Foundation Model for Dexterous Tasks: Laundry Folding

Training a Foundation Model for Dexterous Tasks: Laundry Folding

10:38Improving Laundry Folding with Vision Language Models and Transfer Learning

Improving Laundry Folding with Vision Language Models and Transfer Learning

17:34Generalizing to Unseen Environments with Diverse Data

Generalizing to Unseen Environments with Diverse Data

25:10Responding to Open-Ended Prompts and Interjections with Synthetic Data

Responding to Open-Ended Prompts and Interjections with Synthetic Data

31:17Q&A: High-Quality Data, Funding, Infrastructure, and Model Size

Q&A: High-Quality Data, Funding, Infrastructure, and Model Size