Code as Policies
Translates natural-language robot commands into hierarchical Python policy code that calls deterministic perception and motion-control primitives.
The loop
- Proposes
A code-generation LLM recursively synthesizes Python policy functions and helper routines from a user instruction.
- Judged by
A sandboxed Python interpreter, NumPy/Shapely spatial geometry calculators, and low-level robot impedance controllers execute the policy.
- Steers
Closed-loop visual and geometric state queries inside the generated control loops adjust robot trajectories online.
Headline result
Executed spatial-geometric manipulation, drawing, and tabletop rearrangement tasks on physical Franka Panda and UR5e robot arms without task-specific training.
What this does not prove
Assumes that the perception primitives (object detectors and pose estimators) accurately ground the variable names referenced in the generated code.
How it generalizes the harness
The LLM compiles instructions into executable geometric control code whose loops and spatial math run deterministically on the robot.
Classification
| What it proposes | Program |
|---|---|
| Generative model | Codex / PaLM code-generation model |
| Algorithmic authority | Shapely computational geometry library and classical robot motion controller |
| Feedback | Test results |
| What gets updated | Prompt context |
| When it runs | At inference time |
| Validity | Empirical (measured) |
| Evidence | Experiments reported |
| Lineage | Original system |