Unitree UnifoLM-WLA-1.0: What’s Released and What’s Still Missing

Unitree’s UnifoLM-WLA-1.0 aims to connect scene understanding, predictions about what will change, and whole-body robot action. The distinction behind its open-source announcement is practical: researchers have access to parts of that system, but the provided release-status reports do not establish a complete public release of the roughly 6B-parameter robot-control model.5813

What has Unitree released?

Unitree announced the UnifoLM-WLA-1.0 project and released weights for UnifoLM-ER-1, its roughly 4B-parameter embodied-reasoning component, and UnifoLM-ER-Flow, which incorporates predicted changes and action representations. A September 16 review of the project’s release checklist also marked the UniBot-V1 Challenge Dataset as released.817

That checklist still listed UnifoLM-WLA-Base weights, post-training code, the Unitree-WBT-Dataset, and the Unitree-Manipulation-Dataset as forthcoming. September 20 coverage likewise described the full model weights, code, and remaining datasets as planned rather than available. The released component weights should therefore not be confused with a downloadable, complete 6B whole-body control checkpoint.81320

How the 64-task system is meant to work

Unitree says it trained the WLA system on approximately 2,500 hours of real-robot data and used a single model across 64 tasks: 54 tabletop tasks and 10 whole-body or mobile-manipulation tasks. Its reported configurations include two-finger grippers and multiple five-finger dexterous hands. These are claims about the tasks and hardware represented in Unitree’s evaluation, not evidence that the model works unchanged on any robot.5613

The components have distinct roles. ER-1-4B handles embodied visual and spatial reasoning. ER-Flow adds representations of anticipated scene changes and actions. The proposed WLA system pairs those representations with a flow-based action expert that generates movement, linking an instruction and observed scene to hand, arm, and body control.81730

To coordinate those movements, the reported action-tokenization method encodes continuous end-effector, hand-joint, and lower-body motion as discrete tokens aligned to shared time steps. The point of that alignment is to represent what the robot might change alongside when its different parts should move. It describes the model’s design; it does not, by itself, demonstrate recovery from unexpected events.1822

What do the benchmark results prove?

Unitree reports that ER-1-4B led the compared open models on seven of 16 multimodal perception-and-understanding benchmarks. Those results concern the reasoning component and the models included in that comparison. They are not seven measures of WLA’s end-to-end task success, nor evidence of sustained safety or reliability in deployment.1318

The 64-task demonstrations show the breadth of Unitree’s reported evaluation, but the provided sources do not establish independent reproduction of its real-world performance. The missing full control weights and post-training assets also limit what outside teams can verify. Claims of reliable operation or straightforward transfer to non-Unitree robots should wait for a more complete release, disclosed evaluation procedures, and independent tests across different bodies, sensors, and hands.81322