Stanford researchers introduce RoboArchitect to rewrite robot control code without retraining

RoboArchitect uses a coding model to inspect robot attempts and rewrite the control code around an unchanged robot policy. The code decides when to call the policy, use vision tools or command the arm directly. The researchers report that average success on unseen tasks rose from 12.7% to 40.2% across several benchmarks and policies, without updating model weights. The revised code can also retain values such as effective grasp depths across episodes.