Unitree Robotics has released UnifoLM-WLA-1.0, a general-purpose foundation model designed for humanoid robots. The single 6B-parameter model is trained on approximately 2,500 hours of real robot data and handles 64 distinct tasks, including 10 whole-body actions and 54 tabletop operations.
The architecture builds upon the UnifoLM-ER-1 embodied reasoner based on Qwen3-VL, incorporating future dynamic region prediction via optical flow and VQ-VAE. It discretizes actions using residual VQ for end-effector, hand, and lower body control, topped with an MMDiT action expert for continuous control. The model supports parallel grippers and two different dexterous hands.
Demonstrations on the Unitree G1 show the model performing complex activities such as making beds, loading washing machines, folding clothes, and sorting objects.