Yifei Yao

How Far Does a Frontier Agent Get as an Embodied Policy?

Web report 2026 · Web report

Yifei Yao^*, Yude Zou^*, Guanbo Jiao^*, Ning Gao, Yukai Wang, Jiantong Chen, Zihou Zhu, Xinyu Li, Ningyu Yan, Shuai Wang, Jianming Ma, Haoyu Guo, Xing Gao, Weinan Zhang, Chunhua Shen

* Equal contribution (co-first authors).

Report ↗ Benchmark ↗

Evaluating a frontier agent as an embodied policy through single-shot in-context learning, direct robot control, and analysis of manipulation precision, recovery, and task composition.

An EBench gear-installation observation used in the in-context learning demonstration — A gear-installation observation from the report’s in-context demonstration gallery.

This interactive report evaluates GPT-6-Astra as an embodied control policy on EBench. It compares task-level performance with VLA and world-action policies and examines in-context learning, task composition, recovery, and precision-sensitive manipulation through recorded executions.

The public web report contains the evaluation protocol, results, and case studies. Its linked arXiv technical report is marked as coming soon; the available web report is linked above.

Reference

BibTeX

@misc{yao2026ebenchastra,
  title = {How Far Does a Frontier Agent Get as an Embodied Policy?},
  author = {Yifei Yao and Yude Zou and Guanbo Jiao and Ning Gao and Yukai Wang and Jiantong Chen and Zihou Zhu and Xinyu Li and Ningyu Yan and Shuai Wang and Jianming Ma and Haoyu Guo and Xing Gao and Weinan Zhang and Chunhua Shen},
  year = {2026},
  howpublished = {Online technical report},
  url = {https://internrobotics.github.io/EBench-Astra-Tech-Report/}
}

Metadata checked October 27, 2026. Original source ↗

← All publications