Unitree open-sources UnifoLM-WLA humanoid foundation model

Open-sourcing whole-body humanoid VLA models bridges the gap between high-level vision-language reasoning and low-level physical dexterity, accelerating robotics research beyond proprietary ecosystems. Unified action spaces provide cross-embodiment versatility by allowing a single policy to switch between parallel grippers and five-finger dexterous hands without model architecture rewrites. Furthermore, dynamic region prediction enables interaction-centric world modeling so the policy anticipates scene changes before generating trajectories, while 2,500 hours of real-robot demonstration data drastically mitigates common sim-to-real transfer failures.