China Merchants Lionrock AI Lab

Robot × Agent — Bringing intelligence to machines and warmth to people.

Website · GitHub · Models & Datasets

We are China Merchants Lionrock AI Lab, part of the China Merchants Research Institute of Advanced Technology. We work toward intelligent machines that understand their surroundings, learn through interaction, and help people in everyday life.

Our Robot × Agent vision brings together the reasoning and learning capabilities of general-purpose agents with robotic perception and action.

Research focus

• Embodied AI and robot learning. We study robot foundation models, with a focus on vision-language-action (VLA) models for robotic manipulation. Our research covers deformable object manipulation, bimanual manipulation, and sim-to-real transfer.
• General-purpose agents. We study reinforcement learning for general-purpose agents, focusing on multi-turn tool use and code-as-action interaction with environments.
• World models. We study action-conditioned world models that predict future observations, with the aim of supporting long-horizon robotic manipulation.

Competition highlights

• LeHome Challenge · ICRA 2026 — 1st place, Real-World Challenge. Our joint team sZs demonstrated cloth manipulation and sim-to-real transfer on real robots.
• WorldArena 2.0 — 5th on the Track 1 leaderboard. ActiFlow explores action-conditioned world models, evaluated for simulator video quality. Leaderboard as of October 28, 2026.
• GOAI 2026 — Champion, General Dual-arm Collaborative Manipulation Capability Test. Our GOLION project won this competition in the Embodied Future track.

Explore our work

• L0 — Learning to be an agent. An end-to-end reinforcement learning training pipeline, the NB-Agent code-as-action scaffold, and the L0 model family.
• LiOS — Connecting models to robots. Open infrastructure components for VLA training and cloud–edge robot inference, with low-latency image streaming.
• LeFold — Learning from real manipulation. A preview release of a bimanual cloth-folding dataset in LeRobot format.

Our supporting tools include Etha for model-weight synchronization between training and inference, Scribe for robot dataset visualization and annotation, and Citadel for robot data processing and management.

We share models, data, and code so others can reproduce experiments, explore new ideas, and build on our work. Researchers, developers, and collaborators are welcome.

More information

Learn more about our research and robot demonstrations (articles in Chinese):

• 2025-07-01 · L0: reinforcement learning for general-purpose agents
• 2026-06-02 · LiOS: connecting cloud intelligence with real-world robots
• 2026-08-20 · Lionrock at WRC 2026: real-world cloth manipulation