Chelsea Finn: Building Robots That Can Do Anything

TL;DR

Physical Intelligence is developing a general-purpose robot foundation model (Pi Zero) to replace the current paradigm of building a separate company for every robotics application. They demonstrate that a recipe of pretraining on diverse robot data followed by fine-tuning on curated, high-quality demonstrations—inspired by language model training—unlocks complex dexterous tasks like folding laund

Key insights

• The current robotics approach requires building an entire company from scratch for each application (hardware, software, movement primitives, edge cases), which has historically led to low success in
• Scale from industrial automation lacks behavioral diversity; YouTube data has an embodiment gap; simulation data lacks realism—scale is necessary but not sufficient for open-world generalization.
• For laundry folding, 2-3 months of failure occurred until the breakthrough: pretraining on all robot data then fine-tuning on a curated, consistent, high-quality demonstration set—inspired by language