Building the “App Store” for Robots: Hugging Face’s Thomas Wolf on Physical AI

Thomas Wolf, co-founder and Chief Science Officer of Hugging Face, brought the same prescient vision that led to Hugging Face’s early investment in transformers to robotics through the company’s LeRobot project. He emphasizes democratizing robotics development through open-source tools, diverse hardware approaches and community-driven innovation—mirroring the successful formula that made Hugging Face the largest open-source AI community.

Building communities unlocks exponential growth: Hugging Face’s success in robotics mirrors their transformer strategy—creating accessible tools that transform niche specialists into a broad horizontal community. Their robotics community has grown exponentially to 10,000 developers, proving that providing simple Python-based tools can democratize complex fields and enable software developers to become roboticists.

Diverse form factors beat expensive humanoids: Rather than pursuing costly humanoids that could price out most users, Wolfe advocates for a “galaxy of different form factors” starting with affordable options like their $300 robotic arms. This approach prioritizes accessibility and enables more experimentation, avoiding the elite-only scenario where only wealthy users can afford $100,000 humanoid robots.

Data diversity matters more than data volume: Unlike LLMs that benefit from massive internet-scale datasets, robotics requires diverse, multi-location data to achieve generalization. The key bottleneck isn’t just collecting more robotic task demonstrations, but ensuring sufficient environmental and contextual diversity so robots can adapt beyond their training environments.

Local deployment drives safety and reliability: Robotics demands local model execution more than other AI applications because physical robots losing internet connectivity could cause dangerous failures. This safety imperative makes open-source models particularly valuable in robotics, where running models “as close as possible to the hardware” prevents catastrophic scenarios.

Open science accelerates innovation beyond model sharing: True advancement requires teaching people to train models, not just providing pre-trained weights. Wolfe’s background struggling to access Soviet superconducting research shaped his belief that sharing training methodologies, datasets and implementation details creates more value than releasing models alone—enabling others to build upon and improve the work.