Stanford's RoboArchitect uses a coding model to rewrite robot control code

Stanford researchers present RoboArchitect, a method in which a coding model reviews robot attempts, diagnoses failures, and rewrites the control code over up to 5 rounds without updating model weights.

The paper reports that average success on unseen scenes rises from 9.7% to 43.3%, and that an average run costs under $150. The figures come from Rohan Paul's report of the paper, and all tests ran in simulation. On VLABench, the wrapped general policy outperformed the same policy finetuned on that benchmark.