A robot can understand an instruction yet lack a prebuilt skill that implements it. The key move here is to let generated code express the missing control logic: language selects the behavior, while Python supplies feedback loops, numerical operations, and reusable functions.
Key Findings
- Few-shot examples connect instructions to perception and control APIs, allowing new commands to change both action sequences and policy parameters.
- Recursive generation of undefined functions makes complex policies compositional; the same technique reaches 39.8% pass@1 on HumanEval.
- Demonstrations span tabletop manipulation, drawing, and mobile robots, showing that the representation can work across different action interfaces.
Who Should Read It
A useful starting point for robotics researchers exploring language-conditioned control and program synthesis. Adapting it requires available perception and control APIs; the paper does not establish reliable open-world operation. Quantitative robot evaluation covers constrained simulated tasks, while the physical robot results demonstrate capabilities.