Definition
Plain language
A robot-control style where an AI writes a short program to do a task, runs it, and rewrites it if it fails.
As stated in the literature
An agent design in which a language model generates executable code calling perception, grasping, and motion tools, then debugs that code from execution feedback; successes can be saved as named, reusable, transferable functions.
Also called: Code-as-Policies, Code as Policy
Why it matters: Writing and saving task code as reusable functions lets a robot build up a library of skills it can recombine instead of relearning each job from scratch.
For example, told to stack two blocks, the AI writes a short program that grabs and places each one, runs it, and rewrites the part that fails.
Heard on the show
“Rather than one giant neural network that takes camera pixels in one end and pushes motor commands out the other, this is what's called a Code-as-Policy agent.”Episode 161 — A Robot That Plays Before You Give It a Job, And Why That Beats Retrying