SayCan

SayCan ('Do As I Can, Not As I Say'), published by Google Research in April 2022, was the first major demonstration that large language models could control real robots. The insight: an LLM proposes candidate actions in natural language, but each action is scored by a learned affordance function — a value model trained in the real world estimating whether the robot can actually execute that action right now. Multiplying language probability by affordance probability selects skills that are both useful and feasible.

On a mobile manipulator doing kitchen tasks, SayCan achieved 84% planning success and 74% end-to-end execution success across 101 real-world tasks. It launched the LLM-robotics field and directly led to PaLM-SayCan, PaLM-E, and the broader VLA movement.

Limitations are structural: it requires hundreds of pre-trained individual skills, and the LLM can only compose what already exists. The research code is public in the google-research repository, but SayCan is a method, not a downloadable model.