Gasgoo Munich-On September 17, 2026, Leju Robot officially released the LET2.0 multi-robot consistency real-machine dataset. Built on a production line capable of 10,000 humanoid robots and a scaled training ground, this dataset marks the first time real-machine data from identical robots can be merged for training and reuse. It shifts the paradigm from "single-machine, single-use" to "multi-machine reuse." The initial release includes 2,000 entries of KUAVO real-machine data.
Image Credit: Leju Robot
To tackle data consistency issues caused by assembly tolerances, joint zero positions, and camera placement, LET2.0 establishes a unified calibration, collection standard, and quality control traceability system to merge data across multiple units. The initial batch covers two industrial tasks: picking up medium-sized express bags and sorting switch workpieces. In cross-robot validation, two calibrated KUAVO 5W robots used the same model for parcel sorting, achieving success rates of 88.9% and 86.1%, respectively. This confirms the feasibility of reusing data across different robot bodies.
The LET2.0 dataset spans scenarios including home service, commercial service, industrial operations, and warehousing and logistics. It encompasses three data categories: basic manipulation, whole-body motion control, and force and tactile sensing. Each entry fully records posture, trajectory, visual, and depth information. During collection and annotation, the dataset defines unified requirements for task scenarios, device calibration, and action execution.
Developers can now access the data and toolchain through the OpenLET community. The platform supports the full workflow—from data application and training to deployment and evaluation—creating a closed loop for data and models. This lays the groundwork for data iteration and the growth of the community ecosystem.
