Xingyuan Intelligence Launches Cross-Embodiment Data Engine Chorus Zhonghang to Explore Robot 'Lexemes' and System Scaling
“行元智能”杨世琦：不做本体，押注机器人“词元” | 首发
On October 27, Xingyuan Intelligence launched the cross-embodiment data engine Chorus Zhonghang, aiming to clean and adapt human operation videos for various robots, addressing the issue of insufficient training data for robots. The engine uses technologies such as atomic action recognition and 3D reconstruction to convert human experience into reusable robot training data, achieving approximately 40 times data expansion. Xingyuan Intelligence focuses on data and model system scaling rather than hardware embodiments, enhancing robot task success rates and iteration capabilities in real-world scenarios.
The article highlights the development of 'Chorus Zhonghang', a cross-embodiment data engine that enables human demonstrations to be adapted for multiple robot platforms. It emphasizes the importance of data scaling and system-level integration over hardware-centric approaches, positioning the tool as a foundational component for robot learning and deployment.
Source: Chinese robotics — Robot learning research · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
human demonstrations
40
data scaling factor
40
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
把人类数据 “ 扩容40倍 ” 机器人训练数据短缺是行业老生常谈的痛点。
Open source S18
”杨世琦认为，直接利用人类视频是破解数据瓶颈的必然方向。
Open source S21
人类视频存在手部遮挡、视角变化以及轨迹不完整等问题，加之在身体结构、运动方式等存在差异，若未经处理，绝大多数据会在质量筛选中被直接剔除。
Open source S23
早期依赖遥操作采集数据，成本高昂且场景受限；后来的UMI等方案虽脱离了机器人本体限制，但仍未能彻底盘活现实世界中海量、持续产生的人类行为数据。
Open source S19
行元智能推出的Chorus引擎，核心是要解决机器人学习中的数据规模化问题：如何让有限的人类经验，转化为大规模复用的机器人训练数据。
Open source S24
Source：Chinese robotics — Robot learning research · view.inews.qq.com