Training Data Building the "App Store" for Robots: Hugging Face's Thomas Wolf on Physical AI
238 snips
Sep 9, 2025 
In this engaging discussion, Thomas Wolf, Co-founder and Chief Science Officer of Hugging Face, dives into the world of robotics. He shares his vision for LeRobot, aiming to democratize robotics with open-source tools and affordable hardware. Thomas emphasizes the shift in community-driven innovation, likening today’s robotics landscape to earlier advancements in AI. He explores challenges like data scarcity and the potential to empower software developers to become roboticists. The talk highlights the excitement around transparency and accessibility in advancing technology.
AI Snips
Chapters
Transcript
Episode notes
World Models Unlock Simulated Robot Data
- Advances in image and video generative models enable controllable world models that can simulate diverse scenes.
- Generated simulation data offers a new route to scale robotic training beyond real-world recording.
Favor Diverse, Cheaper Form Factors Over Humanoids
- Humanoids are costly due to many actuators and may not be the optimal first consumer form factor.
- A diversity of cheaper, task-focused robots could drive broader adoption and more use cases.
Mix Local Models With Large Remote Models
- Use both large foundation models and smaller local models depending on latency, cost, and task complexity.
- Route simple or safety-critical behaviors locally and reserve big models for long-horizon reasoning.