-
The Harness: One Agent Harness for Anything that Acts
-
Sihan Xu and Zilin Wang and Hui Shen and Zhengjie Xu and Weiyang Jin and Yuwei Niu and Ye Li and Jiechen Zhang and Yinpei Dai and Yunhao Luo and Wenhao Chai and Xuweiyi Chen and Ziwei Liu and Stella X. Yu
-
Conference on Robot Learning Late Breaking Results, Austin, Texas, 9-12 November 2026
-
Paper
|
Poster
-
Abstract
-
Vision-language-action (VLA) policies are skilled but narrow, while frontier models asked for every action are general but slow and unsafe. The Harness places a coding agent at the top of a three-layer control hierarchy: it writes programs over perception and motion skills, calls served VLAs as bounded skills, and writes guards that react at every control tick without a model call, above a fixed safety shield. Because language models, learned policies, and embodiments enter through configuration, the same harness also operates desktop computers and plays commercial games, with GUI action models in place of VLAs. On the full RoboDojo protocol, one low-cost language model raises the Simate-beta policy it calls from 33.95 to 39.59, at only 3.3 model calls and 0.006 per episode. Videos and more results can be found at https://harness.lmms-lab.com/
-
Keywords
-
agentic robotics
|