Auriga Technical Report: A Task-Agnostic Agent Harness for Zero-Shot Embodied Manipulation

Auriga Technical Report. A Task-Agnostic Agent Harness for Zero-Shot Embodied Manipulation. By Chengcheng Wang and twenty others, from StellarEdge AI. Published on the thirty-first of October.

Vision-language models can interpret instructions and plan manipulation tasks. Where they struggle is acting directly as robot policies. On RoboDojo, the report measures a large gap between planning and execution, separating the high-level task decisions from what physically happens. One number makes it concrete. With the reference control interface, GPT-6 Astra succeeds twenty-two point five percent of the time. Remove arm execution from the picture, and its success rate rises to seventy-six point two percent. Much of the failure, in other words, lives in the acting rather than the deciding.

Auriga is a task-agnostic agent harness built to close that gap, and it closes sixty-eight point three percent of it. The setting is zero-shot manipulation, with no trials on the target task and no resets. The division of labour is clear. The model makes all of the task decisions, reasoning over an explicit task belief. The harness supplies what the model is poor at providing for itself.

The model acts in one of two ways: direct tool calls, or synthesized Python programs that run several steps. The harness, for its part, contributes calibrated 3D measurements, dual-arm control referenced to the fingers, and a single admission gateway through which everything reaches the robot. Checks happen on both sides of motion. Before moving, the model inspects a visual preview of its trajectory while the simulation is paused. During execution, the harness checks declared stage tolerances and grasp bounds. If the robot diverges, the harness stops it and returns the measured discrepancy, so the model can recover in place rather than starting again.

The harness was developed across thirteen cross-task capability iterations, then frozen for evaluation. On the RoboDojo leaderboard it ranks first, with a fifty-seven point seven percent success rate and a Score of sixty-four, more than doubling the reference harness on the same model. Under scene randomization it stays first, while learned policies drop by up to thirty-three point nine percent.

The same frozen harness was then taken to RoboCasa365. There it reaches a seventy-three point two percent success rate, and on composite tasks it outperforms agents assisted by VLA models by up to twenty-three point one percent.

Nothing about the underlying model changes between the reference interface and Auriga. What changes is the scaffolding around it: measurements, a preview before motion, tolerance checks during motion, and a way back when something goes wrong. The doubling on the same model is the clearest sign of how much of that original gap belonged to execution rather than to judgment.