Leju CEO: Real-Robot Data Is Key to Scaling Embodied AI

At the 2026 Embodied Intelligence Robot Conference held on October 28, Leju CEO Chang Lin said that real-robot data is central to scaling embodied AI applications. He outlined a training pipeline consisting of pretraining, mid-training, post-training, and reinforcement learning on physical robots, with each stage requiring different data types.

Chang noted that pretraining builds general-purpose models using first-person view data, hand-held gripper collection, and real-robot data. For specific industries, platform companies train vertical domain models with thousands of hours of teleoperation data across three to five scenarios, after which developers perform post-training with about 100 to 200 hours of teleoperation data before collecting in-loop data for reinforcement on real machines.

Data reuse, he said, determines the ability to scale deployments. Leju has already achieved routine operation of about 15 robots in an auto parts machining and loading scenario at Hangzhou Zhaofeng Electromechanical, with efficiency reaching 60% to 70% of human workers.

Leju also leads the OpenLET developer community to connect domestic data, models, and computing resources. It has open-sourced 100,000 minutes of real-robot data, with downloads exceeding 6 million, covering first-person view, simulation, and synthetic data. Chang said open data and toolchains lower development barriers and encourage broader participation in embodied AI applications.