Leju Robotics has unveiled LET2.0, a multi-machine consistency dataset for humanoid robots, marking a step toward reusing real-world robot data across multiple units rather than confining it to a single machine, according to Chinese outlet Gasgoo.
Announced on September 17, 2026, the dataset was built using a production line capable of turning out tens of thousands of humanoid robots and a large-scale training facility. Leju says it is the first time that real-world data from identical robot models has been merged across machines for combined training and reuse, shifting the paradigm from "one machine, one use" to "multi-machine reuse." The initial release includes 2,000 entries of KUAVO real-world data.
The core challenge addressed by LET2.0 is data inconsistency caused by assembly tolerances, joint zero-position offsets and camera position differences between units of the same model. To solve this, Leju established unified calibration, collection standards and a quality inspection and traceability system, allowing data from different machines to be merged into a single training set.
The first batch of data covers two industrial tasks: picking up medium-sized express bags and sorting switched workpieces. In cross-machine validation tests, two calibrated KUAVO 5W robots using the same model achieved package-sorting success rates of 88.9% and 86.1%, respectively, demonstrating the feasibility of cross-body reuse, the company said.
Dataset scope and developer access
LET2.0 covers scenarios including home services, commercial services, industrial operations and warehouse logistics. It contains three types of data: basic manipulation, whole-body motion control and force/tactile sensing. Each data entry records posture, trajectory, visual and depth information. During collection and annotation, the dataset specifies unified requirements for task scenarios, equipment calibration and action execution.
Developers can now access the data and toolchains through the OpenLET community. The platform supports a complete workflow from data application, training and deployment to evaluation, creating a closed loop between data and models, which Leju says lays the groundwork for data iteration and community ecosystem development.
The release comes as Chinese humanoid robot developers race to gather high-quality training data, with data scarcity seen as a key bottleneck for advancing robot foundation models. Leju's approach of standardizing data across multiple machines could help lower the cost of data collection for industrial deployment.
