Code as Policies: the robot programs itself from natural language
In one sentence Google shows how an LLM directly generates executable robot code from natural-language instructions, without robotic fine-tuning, using hierarchical function composition.
Instead of training a separate model for every type of robot task, Code as Policies uses a general-purpose LLM to generate Python programs that the robot executes directly. The instruction "put the red cube on top of the blue one" becomes code with calls to vision, manipulation, and control functions.
The key point is hierarchical composition: the LLM can write functions that call other functions, building complex solutions from simple available building blocks. No robotics-specific fine-tuning needed.
The approach works on both manipulation robots and drones, and generalizes to new instructions never seen during development.
Companies
Tools
Code as Policies, PaLM, Codex
Tags
Sources