Unitree UnifoLM-WLA-1.0: what is actually open now?

The most consequential UnifoLM-WLA-1.0 update since its announcement is not another robot video. On 20 September, Unitree said it released model modules and code for training an action expert. The repository now supplies a training guide and a processing specification; the UnifoLM-ER-1 and ER-Flow weights had already been released on 11 September. That is a more usable research starting point than a project page alone, but it is not the same thing as releasing the complete model shown in the demonstrations. [S1] [S2] [S3]

The distinction matters because “open source” describes several different things here: a vision-language backbone, an action-generating component, training code, robot data and a trained whole-body policy. A researcher can work with some of those pieces today. Others remain on Unitree’s release plan.

What a developer can obtain

Unitree links to downloadable UnifoLM-ER-1 and UnifoLM-ER-Flow model repositories. Their model cards identify them as 4B-parameter models. ER-1 is the embodied-reasoning vision-language model; ER-Flow adds representations for predicted scene changes and discrete actions. Neither weight set should be described as the advertised 6B UnifoLM-WLA-1.0 whole-body policy. [S1] [S4] [S5]

The new guide describes a concrete path to train an action expert from scratch on top of either ER model. It tells users to point the launch script at a local backbone, arrange Dex1 and whole-body-task data, edit the multi-source dataset configuration and run the MMDiT training script. The repository includes Python packaging and a lockfile. Its guide says the script uses the GPUs detected on a node by default; this is research training code that requires substantial local data, compute and configuration, not a one-click robot application. PHASE has inspected the public instructions but has not run the training or tested a trained policy on hardware. [S1] [S2]

One useful technical release is the action-and-state specification. It defines a 54-dimensional unified action vector, a 60-dimensional robot state, validity masks for absent modules and a 120-dimensional state-projector input formed from state plus mask. It also fixes coordinate, normalization and sampling conventions. Those details make cross-dataset training less ambiguous; they do not by themselves show that a resulting controller works across hands, tasks or robots. [S3]

What is still missing

The repository lists UnifoLM-WLA-Base, code for fine-tuning that base model and LoRA fine-tuning as planned items, without corresponding released links. Unitree’s WLA collection currently points to the ER weights and dataset collections, not a WLA-Base checkpoint. A publicly downloadable pretrained WLA action expert or an end-to-end deployment recipe for reproducing the project-page G1 demonstrations is not identified in the current repository documentation. This is a statement about the material PHASE could verify on 26 September, not a claim that Unitree has no internal model or that further releases will not follow. [S1] [S2] [S6]

The dataset links are meaningful, but their existence does not establish that the full training mixture, selection, preprocessing and evaluation setup behind the advertised system is public. The project page says the WLA model was trained on approximately 2,500 hours of real-robot data and performs 64 tasks, including ten whole-body and 54 tabletop tasks. It presents robot-test videos and describes an ER-Flow backbone with an MMDiT action expert. These are Unitree’s claims and demonstrations; PHASE has not independently verified the hours, task count, success rates or transfer to a different setup. [S1] [S6] [S7]

The practical reading is therefore neither “just an announcement” nor “the complete humanoid policy is open.” The September releases give researchers weights for the ER backbones, model components, training code, data links and unusually explicit action/state conventions. The decisive missing comparison is still a released WLA-Base checkpoint, accompanied by enough inference, deployment and evaluation detail for outsiders to reproduce the claimed whole-body behaviour. Until then, the public stack is a route to build and test an action expert, not the finished 64-task system in a download. [S1] [S2] [S3] [S6]