FAST tokenizer

Chelsea Finn describes Physical Intelligence’s effort to build a general-purpose model that lets different robots perform many tasks in varied settings. She explains how large-scale real robot data, pre-training, and fine-tuning on curated demonstrations helped robots handle laundry and other dexterous tasks. Diverse training data and improved language following helped the robots work in unfamiliar homes, while synthetic prompts enabled more open-ended interactions. The talk also covers remaining challenges in reliability, speed, planning, and real-world deployment.