RoboArchitect rewrites robot control code automatically
The key result is that test-time control code rewriting can beat finetuning itself, so adapting a robot policy no longer requires collecting new demonstrations or touching weights. Robotics engineers and embodied AI teams can treat it as a cheap wrapper around existing general policies, with the caveat that every reported number came from simulation only. That simulation-only gap is the main open question before real hardware adoption.
A new Stanford paper introduces RoboArchitect, a system where a coding model reviews recorded robot attempts, diagnoses failures, and rewrites control code over up to 5 rounds, with no model weights updated. An average full run cost under US$150. On unseen scenes, success rates rose from 9.7% with the bare policy to 43.3%, and on unseen tasks from 12.7% to 40.2%. On VLABench, the wrapped general policy outperformed the same policy finetuned on that benchmark, though all tests ran in simulation.
New Stanford paper shows that robot control code no longer needs hand-building for each new robot, because a coding model can diagnose failures and rewrite it automatically.
Robot policies often fail on new setups, and fixing them usually means new demonstrations and finetuning. Wrapping the policy in control code helps, but engineers write it by hand for each robot.
RoboArchitect has a coding model review recorded robot attempts, diagnose failures, and rewrite the control code over up to 5 rounds. A full run cost under $150 on average, with no weights updated.
On unseen scenes, average success rose from 9.7% for the bare policy to 43.3%, and on unseen tasks from 12.7% to 40.2%. On VLABench, the wrapped general policy beat the same policy finetuned on that benchmark, though all tests ran in simulation.
