Generalist, a robotics AI startup training foundation models across multiple robot types and tasks, has raised $400 million in new funding. The company aims to build general-purpose robot intelligence that can outperform task-specific systems, similar to how large language models succeeded in software. The startup focuses on solving robotic dexterity, which founders consider the hardest unsolved problem in robotics and the clearest path to commercial value. Generalist was founded by Pete Florence, former DeepMind senior scientist and author of influential papers including PaLM-E and RT-2, alongside Chief Scientist Andy Zeng and CTO Andrew Barry from Boston Dynamics. The founding team combines researchers from DeepMind, OpenAI and Boston Dynamics. Early results show rapid adaptation to new robots and tasks with strong sample efficiency and commercial traction.

Generalist has announced GEN-1, a physical AI system that achieves 99 per cent success rates on delicate mechanical tasks like folding boxes and packing phones, operating roughly three times faster than its predecessor. The model can respond to disruptions by improvising new moves, such as shaking a bag to position objects or refolding shirts that shift mid-task. It reaches production-level performance after just one hour of adaptation to specific robotic hardware. Generalist trained the system using over 500,000 hours of data captured through wearable "data hands" that record human manipulation of objects. The company says GEN-1 represents a GPT-3-style inflection point for physical robotics, where models can now perform economically useful work across increasingly complex tasks without explicit programming for error recovery.

Generalist Bets Wearable “Data Hands” Could Spark Robotics’ ChatGPT Moment