Unitree open-sources UnifoLM-WLA-1.0, a 6B whole-body VLA: 64 tasks, Apache 2.0 — and a gripper-only inference server
A Qwen3-VL-based reasoner, RVQ-discretized actions, and an MMDiT flow expert with LeRobot datasets — plus self-reported benchmarks and a whole-body gap in what actually ships.
Unitree has open-sourced UnifoLM-WLA-1.0, which it describes as a ~6B vision-language-action (VLA) model running 64 whole-body and tabletop manipulation tasks on the G1 humanoid from a single checkpoint. Weights, fine-tuning code, and LeRobot-format datasets ship under Apache 2.0. The release targets whole-body humanoid control rather than tabletop-only manipulation — but the shipped artifacts don't quite match the demo reel.
What shipped
Unitree announced UnifoLM-WLA-1.0 on September 10, 2026, then released it in stages. The embodied-reasoner weights (UnifoLM-ER-1 and UnifoLM-ER-Flow) landed on Hugging Face on September 11; model modules and action-expert training code followed on September 20; the full UnifoLM-WLA-1.0-Base checkpoint and fine-tuning code arrived on September 28. All three model repos and the GitHub code repo are Apache 2.0.
The architecture is a three-part stack, detailed on the project page:
- UnifoLM-ER-1, the embodied reasoner, is built on Qwen3-VL-4B. Unitree says it trained on more than 5 million samples spanning image point prediction, object detection, multi-image reasoning, 2D trajectory prediction, 3D object detection, and multi-image spatial question answering, co-trained with general image-text data to preserve broad vision-language ability.
- Future dynamic region prediction: optical flow extracts where the scene will change, a VQ-VAE encodes those changes into fixed-length discrete mask tokens, and the VLM predicts the future dynamic-region tokens conditioned on the current image and task.
- Discrete action learning: a unified action space split into end-effector poses, end-effector joints (gripper or dexterous hand), and lower-body joints, each discretized by a separate residual vector quantization (RVQ) model. The tokens share timesteps and feed synchronously into the VLM.
- UnifoLM-WLA-1.0 adds an MMDiT action expert — a flow-matching decoder — on top of the ER-Flow backbone for continuous control, trained jointly on language, discrete-action, and continuous-action objectives.
The shipped config gives the numbers engineers care about: a 54-dimensional unified action vector, a 60-dimensional state vector, a 120-dimensional robot-state input, 30 Hz control with 30-step action chunks (roughly a one-second horizon), and an MMDiT with 16 layers, hidden size 1024, and 4 inference timesteps. The ER-1 checkpoint is ~4.44B parameters and ER-Flow ~4.45B; the WLA-1.0-Base repo is ~12.5 GB on disk. The G1 itself has 29 degrees of freedom.
Unitree says training used ~2,500 hours of real-robot data, including its own open datasets and the BitRobot-HIW-500 collection; the project page mentions no simulation data. It splits the 64 tasks into 10 whole-body and 54 tabletop, and says the model supports two-finger grippers plus multiple five-finger dexterous hands — the released whole-body datasets name BrainCo and Inspire hands. Demo videos show the G1 making beds, loading a washing machine, folding towels, and sorting small components from conveyor belts.
Fine-tuning is cheap: the VLM is frozen by default, only the DiT action head and robot-state projector train, and the default config targets a single 24 GB GPU. LoRA is supported for both the VLM and the action head.
On the reasoner side, Unitree reports UnifoLM-ER-1 leads open-source models on 7 of 16 spatial and multimodal understanding benchmarks (RefSpatial-Bench, Where2Place, Pixmo-Point, BLINK, EmbSpatial, RoboSpatial, and VSR) and beats its own Qwen3-VL-4B base on 12 of 16, losing on VSI-Bench, RealWorldQA, MME, and MMMU_VAL. On Where2Place it reports 82.0 against GPT-6-Astra's 69.0. All of these numbers are Unitree's own testing.
Why it matters
UnifoLM-WLA-1.0 treats the whole body — legs, waist, arms, and hands — as one action space, and ships the data and code to reproduce the pipeline, not just a checkpoint. That makes it a usable baseline for teams building humanoid control, not another demo video.
The design is a clean, liftable recipe: keep a general VLM as the reasoner, add future-region prediction as an intermediate supervision signal, discretize actions with RVQ so the VLM can emit them as tokens, and bolt on a flow-matching expert for continuous control. That separation — discrete tokens for the VLM, continuous decoding for the policy — is documented here and Apache-2.0.
For practitioners, the win is fine-tuning: a frozen VLM plus a trainable action head that fits a single 24 GB GPU lets a robotics team adapt the model to its own embodiment and data without a cluster. The released whole-body datasets (LeRobot format, 30 fps, four cameras) give a concrete starting point.
Caveats
- Everything is self-reported. Unitree's benchmark table footnotes say BLINK scores average only the Relative Depth and Spatial Relation subtasks, and proprietary-model results came from API tests. No third party has reproduced the numbers; Unitree publishes no real-robot success rates.
- The shipped inference server only covers the two-finger-gripper (Dex1) action space. Serving the whole-body or dexterous-hand checkpoints requires extending the action protocol — the repo's own docs describe the missing entries (
left_fig6d/right_fig6dfinger angles,base_posewhole-body motion). The whole-body demos are not what you get out of the box. - No paper. The repo cites the model with a
@miscentry; there is no arXiv technical report, and no latency spec in the repo. The sub-105 ms inference figure circulating comes from a Unitree telemetry caption in coverage, not from the docs. - "6B" is approximate. The ER-1 backbone is ~4.44B parameters; Unitree doesn't publish a breakdown of how the backbone plus action expert sums to the headline 6B.
- Unitree's own CEO is cautious. Wang Xingxing has said humanoids are 2–10 years from an 80%-success "ChatGPT moment" on voice-prompted tasks in unfamiliar rooms.