Xingyuan Intelligence Releases Cross-Embodiment Data Engine Chorus, Claiming Approximately 40x Training Data Expansion

Xingyuan releases cross-embodiment data engine Chorus: "device-agnostic," enabling roughly 40x expansion of training data. Zhong Lilei, 1 hour ago. Achieves over 40x data expansion.

On October 27, Action Intelligence (Xingyuan Intelligent) officially released "Chorus Zhongxing," a cross-embodiment data engine for robot pre-training. It supports converting human manipulation videos captured by various devices—such as smart glasses, chest-mounted cameras, wrist cameras, UMI devices, and smart headbands—into pre-training data adapted to different robot embodiments, improving the reuse efficiency of human data in robot training. Tests show that Chorus can achieve over 40x expansion of effective robot training data.

Currently, academia and industry are jointly exploring how to use automated data engines to convert human manipulation data into robot training data. However, high collection costs, low usability of raw data, and difficulty in cross-embodiment reuse still constrain the scaling of data supply. Chorus's core approach is to maximize the value of raw data by ingesting human manipulation data from more sources, fully mining the motion information within it, constructing multi-granularity instructions, and adapting to different robot embodiments.

Chorus is a cross-embodiment data engine independently developed by Xingyuan. Built on a unified data model, it connects multi-source human manipulation videos with different robot embodiments through a highly automated data processing pipeline. It has three core capabilities: unified representation of diverse heterogeneous data, robust reconstruction of complex manipulation behaviors, and cross-embodiment motion generalization and augmentation.

First, Chorus supports human manipulation videos captured by various devices, including smart glasses, chest-mounted cameras, wrist cameras, and UMI-type devices. It brings data that meets processing requirements—across different hardware, viewpoints, and collection conditions—into a unified pipeline, reducing dependence on a single collection device and expanding the available data sources.

Second, unlike the industry's common approach of "one dataset for one robot," Chorus can convert a single human demonstration into training data adapted to different robot body structures and motion ranges, achieving cross-embodiment reuse. By sharing the earlier data processing results, when a new target embodiment is added, the system does not need to repeat all source data processing, reducing computational overhead.

During data processing, Chorus uses 3D motion reconstruction, trajectory optimization, and quality verification to reduce the impact of factors such as camera movement and hand occlusion on conversion results, raising the average usable data conversion rate to 85%. Under the current test configuration, Chorus, through adaptation to 16 robot embodiments, reduces the processing cost per hour of effective robot trajectories to 15 yuan; combined with 3x instruction augmentation, it can achieve approximately 40x expansion of effective training data, with the equivalent cost per unit of training sample only 5 yuan/hour.

Action Intelligence is an emerging embodied intelligence company positioned to continuously transform human experience into robot capabilities and build a general embodied brain for the physical world. Xingyuan believes that the scaling of embodied intelligence cannot be measured solely by compute and parameter scale, but also by whether it can be transformed into robots' ability to solve problems in the real world.

To this end, Action Intelligence adopts the System Scaling technical roadmap, connecting data, models, and real-world deployment: the previously released Agent Continuo can be deployed on wearable devices, supporting the scaling of human motion videos; through Chorus, multi-source human manipulation demonstrations are converted into robot pre-training data; and "Duet Tongxing," planned for launch within the year, will accelerate the entry of models into real scenarios, continuously feeding human guidance, corrections, and interaction feedback back into model training, promoting mutual reinforcement between model capabilities and application scale, and forming a closed loop of continuous learning and iteration.

Around this technical roadmap, Action Intelligence has assembled a core R&D team covering data, models, training infrastructure, and robot deployment. Founder and CEO Yang Shiqi graduated from the University of California, San Diego, where he studied under Professor Wang Xiaolong, Research Director at Meta Superintelligence Labs, and is one of the world's earliest researchers in cross-embodiment manipulation and human data learning. Technical co-founder Si Chenyang is a national-level young talent with deep research in foundation model architecture and large-scale model training, and is responsible for the architecture and scaling research of Xingyuan's WAM world action model. Technical co-founder Yang Jingkang is a representative researcher in the field of first-person multimodal perception and human data understanding. Recently, both were selected for the 2026 list of the world's top 2% scientists.

Source: Submission

Images in this article are from: 123RF licensed image library. Infringement report. Back. Share. Related articles. All comments.