Auriga Technical Report: A Task-Agnostic Agent Harness for Zero-Shot Embodied Manipulation
Auriga more than doubled the same model’s robot-task success rate, reaching 57.7% on RoboDojo.
The authors report that GPT-6 Astra’s success rose from 22.5% to 76.2% when arm execution was removed, exposing a gap between planning and physical control. Auriga supplies calibrated 3D measurements, dual-arm controls and a visual trajectory preview before motion. During execution, it stops on deviations and returns measured discrepancies so the model can recover in place. The harness was frozen before evaluation, required no target-task trials or resets, and also achieved 73.2% success on RoboCasa365.
HF Daily Papers' note
The authors report that GPT-6 Astra’s success rose from 22.5% to 76.2% when arm execution was removed, exposing a gap between planning and physical control. Auriga supplies calibrated 3D measurements, dual-arm controls and a visual trajectory preview before motion. During execution, it stops on deviations and returns measured discrepancies so the model can recover in place. The harness was frozen before evaluation, required no target-task trials or resets, and also achieved 73.2% success on RoboCasa365.
HF Daily Papers' note
score 5