Auriga harness tops RoboDojo at 57.7%, doubles baseline

Auriga, a task-agnostic agent harness, ranks first on the RoboDojo leaderboard with a 57.7% success rate (64.0 Score) in zero-shot manipulation, more than doubling the reference harness running the same underlying model, according to the Auriga Technical Report posted to arXiv.

The paper frames the result as evidence of a wide planning-execution gap in vision-language-model robot agents. On RoboDojo, 'removing arm execution raises GPT-6 Astra's success rate from 22.5% under the reference control interface to 76.2%,' the authors write, isolating physical control rather than reasoning as the bottleneck. Auriga 'closes 68.3% of this gap in zero-shot manipulation without target-task trials or resets.'

Architecturally, the model acts over 'an explicit task belief' and reaches the robot through 'either direct tool calls or synthesized multi-step Python programs,' while the harness supplies calibrated 3D measurements, finger-referenced dual-arm control, and a single admission gateway. Before motion the model inspects a visual trajectory preview with simulation paused; during execution the harness checks stage tolerances and grasp bounds, stopping the robot on divergence and returning the measured discrepancy for in-place recovery.

Auriga was developed across 13 cross-task capability iterations and then frozen for evaluation. On RoboCasa365, the same frozen harness reaches a 73.2% success rate, outperforming VLA-assisted agents by up to 23.1% on composite tasks. Under scene randomization on RoboDojo, learned policies drop by up to 33.9%, while Auriga stays on top.