Learning-Deep-Learning

SayCan: Do As I Can, Not As I Say: Grounding Language in Robotic Affordances

July 2023

tl;dr: Generate natural language actions that are both feasible and contextually appropriate. Combine generated action and the affordance to pick the most probable action. Overall score = say * can.

Overall impression

SayCan’s key idea is to ground language models through affordance functions. The affordance functions capture the log likelihood that a particular skill will succeed in a particular state.

Critiques:

Advantages:

This is a quite big project, with many authors contributing to many aspect of this complex engineering project.

Key ideas

Technical details

Notes