Auriga harness doubles GPT-6 Astra zero-shot robot success on RoboDojo
Original title: Auriga Technical Report: A Task-Agnostic Agent Harness for Zero-Shot Embodied Manipulation
The frozen harness posts 57.7% success and stays first under scene randomization while learned policies fall by up to 33.9%.
Auriga is a task-agnostic agent harness meant to close the gap between vision-language planning and physical robot execution. The model decides over an explicit task belief and acts through tool calls or synthesized Python, while the harness supplies 3D measurements, dual-arm control, trajectory preview, and tolerance checks that stop motion on divergence. Frozen after 13 capability iterations, it ranks first on RoboDojo at 57.7% success (64.0 Score) and reaches 73.2% on RoboCasa365.
Unconfirmed: Success rates and leaderboard rank are author-reported; independent replication is not stated.
- Removing arm execution raises GPT-6 Astra from 22.5% to 76.2% on RoboDojo.
- Auriga closes 68.3% of that planning-execution gap without target-task trials.
- It more than doubles the reference harness on the same model.
- On RoboCasa365 it beats VLA-assisted agents by up to 23.1% on composite tasks.
- Products
- Auriga
- AI models
- GPT-6 Astra
From the original article
1 min read at huggingface.co
Vision-language models can interpret instructions and plan manipulation tasks, yet struggle when acting directly as robot policies. On RoboDojo, a large planning-execution gap separates high-level task decisions from physical execution: removing arm execution raises GPT-6 Astra's success rate from 22. 5% under the reference control interface to 76.
By Chengcheng Wang, Yuheng Shi, Siyu Xu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.