China Merchants Lion Rock AI Lab debuts at WRC 2026 with flexible clothing folding tech

Cloth folding is widely recognized as the 'technical touchstone' in the global embodied intelligence field, because clothing is an amorphous flexible object with strong dynamic uncertainties in material thickness, wrinkle patterns, entanglement states, surface friction, and elastic deformation. For robots, this tests five core capabilities: flexible perception, dual-arm coordination, contact force control, long-horizon operations, and state recovery, with difficulty increasing exponentially with clothing type, initial disordered state, and topological complexity.

Current mainstream cloth folding solutions generally suffer from the pain point of 'excellent in simulation, poor on real robots, and limited in scenarios.' Most technologies only work in ideal environments with neatly arranged clothing. In real home scenarios with wrinkled piles, random orientations, partial entanglement, and clumped compression, they are prone to grasp failures, folding misalignments, process interruptions, and task failures. The core issue lies in the sim-to-real gap: hardware rigidity differences, mechanical assembly errors, insufficient gripper stability, and low-level control precision deviations cause smooth simulation actions to drift, grasp instability, and grasp failures on real robots.

To address the sim-to-real challenge, the lab built a data iteration loop covering training, deployment, trajectory sampling, and Real2Sim teleoperation. On the training front, distributed parallel training strategies and high-performance operator optimizations were introduced, improving model training throughput efficiency by more than 5 times. On the evaluation front, the simulation environment was refactored for parallelism, improving overall evaluation efficiency by more than 4 times.

Leveraging its self-developed VLA model and systems engineering capabilities, the team advanced to the finals of the ICRA 2026 LeHome Challenge in June 2026, defeating world-class opponents including the 2025 BEHAVIOR-1K champion and Ilya, and securing first place globally. This validates the leading performance of its technology in real-robot flexible manipulation tasks.