Robotics labs keep reaching for the same lever: a bigger policy, a longer finetune, a newer base model. A Stanford group led by Serena Yeung-Levy just published evidence that a different lever was sitting there the whole time — the harness wrapped around the policy.
The Harness Is Where the Robot Lives
A VLA policy is a frozen mapping from pixels and instructions to motion chunks. It is good at what it saw and brittle everywhere else. The paper’s opening claim is that a frozen policy performs dramatically better inside an agent harness: control code that decides when to query the policy, when to call an auxiliary vision or grasp model, and when to bypass the policy entirely and command the arm directly.
Those harnesses are hand-written today, and rebuilt from scratch for every new robot and benchmark.
RoboArchitect automates the design. A coding model runs the agent on training episodes, records every step (camera frames, robot readings, every model and tool output), diagnoses the failures, and writes a revised harness. The revision replaces the incumbent only if it scores higher on the same episodes. After N iterations the final agent is evaluated on held-out episodes never used for optimization.
What It Bought
Average success rose from 9.7% to 43.3% on unseen scenes (4.5x) and 12.7% to 40.2% on unseen tasks (3.2x), across policies and benchmarks. The per-cell numbers are starker than the averages: OpenVLA-7B on LIBERO-Pro-10 went from 0.0% to 26.0% on unseen scenes, RLinf-pi0.5 trained on LIBERO-130 from 2.0% to 40.0%, SmolVLA on RoboCerebra from 0.0% to 38.0%, and X-VLA on SimplerEnv WidowX from 0.0% to 87.5%.
The most quoted result is a direct comparison with finetuning. X-VLA is a general-purpose policy never trained on VLABench. The evolved harness around the unfinetuned model scored 52.6% on unseen scenes and 72.5% on unseen tasks. X-VLA finetuned on VLABench scored 35.2% and 32.1%. The harness beat the weights. Iterative refinement also beat one-shot agent design on nearly every held-out set.
The harnesses that came out read like something a person wrote: reaches in from the side under a shelf board, reaches the named object, not a memorized place, pushes until the door latches shut. The agent also grows a persistent memory with no fixed schema — the coding model decides the entries, their format, and when the agent reads them — carrying experience like which grasp depth worked across episodes. The same approach produces one unified harness spanning the Franka arm, WidowX and the Google robot.
Everything here is simulation (LIBERO-Pro-10, RoboCerebra, VLABench, RoboCasa-365, SimplerEnv), and the optimisation loop needs rollouts to run on.
Related: arXiv:2610.42141 · Project page