Post by Ed Henderson

We just taught a robot to play Jenga. Not with scripted motions. Not with a rules engine. By showing it, about forty minutes of Benjamin Giudice and I moving two robot arms through the task, and letting a frontier AI model learn what we were doing. This is a big deal for us in our journey. The model is π₀.₅ from Physical Intelligence — a vision-language-action model, the robotics equivalent of what GPT was for text. The arms are SO-101s from Hugging Face's LeRobot project. Is it perfect? No. Our robots are slow. But they finish the task. This is the earliest signs of an important milestone in human history - intelligence in the physical world. The ingredients are finally on the shelf: open-weights models, affordable hardware, cloud compute by the minute, and a community building the tooling to connect it all. #robotics #physicalAI #embodiedAI

Video Content