Understand this story
On October 30, 2026, Xingyuan Intelligence released Chorus (Zhongxing), a cross-embodiment data engine for robot pre-training. According to a Leiphone report, the engine can convert human manipulation data collected by devices such as smart glasses, chest-mounted cameras, wrist cameras, and UMI into training data adapted to different robot embodiments. Xingyuan Intelligence said the average usable data conversion rate has risen to 85%, it adapts to 16 robot embodiments under the current test configuration, and the processing cost per hour of effective robot trajectories has dropped to 15 yuan; combined with 3x instruction augmentation, it can achieve roughly 40x expansion of effective training data. The report also outlined the difficulties in the five stages of embodied data—filtering, annotation, motion conversion, validation, and feedback iteration—as well as the respective shortcomings of the four data sources: real robots, simulation, UMI, and first-person human video. No further progress has been disclosed so far.
AI-generated from public reports
Developments
1 confirmed developments
Xingyuan Intelligence releases cross-embodiment data engine Chorus
Source:English text is being prepared.
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 30
Leiphone
Xingyuan Intelligence Releases Cross-Embodiment Data Engine Chorus, Untangling the Challenges of the Embodied Data Pipeline
Xingyuan Intelligence has released Chorus (Zhongxing), a cross-embodiment data engine for robot pre-training, which can convert human manipulation data collected by devices such as smart glasses, chest-mounted cameras, wrist cameras, and UMI into training data adapted to different robot embodiments. Xingyuan claims that the average usable data conversion rate has risen to 85%, that it adapts to 16 robot embodiments under the current test configuration, and that the processing cost per hour of effective robot trajectories has dropped to 15 yuan; combined with 3x instruction augmentation, this can achieve roughly 40x expansion of effective training data. The article also outlines the difficulties in the five stages of embodied data—filtering, annotation, motion conversion, validation, and feedback iteration—as well as the respective shortcomings of four types of data sources: real robots, simulation, UMI, and human first-person videos.