AY-ROBOTSAI Teleoperation
Sign Up
UNI

UnifoLM-VLA-0

Unitree, China · January 2026

Unitree Robotics

Foundation VLAOpen weights, CC BY-NC-SAAlso written UnifoLM-VLA-Base, UnifoLM-VLA-LIBERO
Parameters
not published
No total parameter count is published for the VLA model.
GPU memory
not published
weights at bf16, computed
Inference latency
not published
per action step
Weights
CC BY-NC-SA
CC BY-NC-SA 4.0

What it is

UnifoLM-VLA-0 is Unitree's open-weight vision language action model for general purpose humanoid manipulation, released on 29 January 2026 without an accompanying arXiv paper. It continues pretraining of a Qwen2.5-VL-7B backbone on 2D detection and segmentation, hierarchical task decomposition, 3D object detection, spatial reasoning and trajectory prediction data, producing the intermediate UnifoLM-VLM checkpoint, then attaches a diffusion transformer action head. The VLA stage uses roughly 340 hours of real robot data recorded on the Unitree G1 humanoid together with action chunking and forward and inverse dynamics constraints. Unitree reports a 98.7 percent LIBERO average, the top entry in their own comparison table. The main caveats are documentation gaps: no total parameter count is published, the spatial understanding comparison exists only as an image without a numeric table, no real robot success rates are given, and the CC BY-NC-SA 4.0 license rules out commercial deployment.

Architecture

Backbone
Qwen2.5-VL-7B, continued pretrained into UnifoLM-VLM (released as UnifoLM-VLM-Base)
Action head
Diffusion transformer action head. The released config.yaml defines action_model.diffusion_model_cfg with 16 layers, hidden size 1024, cross_attention_dim 2048, 32 attention heads, ada_norm, and num_inference_timesteps 4, over a beta noise schedule (noise_beta_alpha 1.5, noise_beta_beta 1.0).
Action chunk
16 steps per forward pass
Parameters
No total parameter count is published for the VLA model. The project page names Qwen2.5-VL-7B as the backbone. The released UnifoLM-VLM-Base repository reports 8,292,166,656 BF16 parameters in its safetensors index (Hugging Face model API, checked 11.08.2026). UnifoLM-VLA-Base ships one checkpoints/pytorch_model.pt blob of 18,982,885,927 bytes and exposes no safetensors parameter index, so the VLA total cannot be derived without guessing.
Pretraining data
Stage 1 (VLM): continued pretraining on a multi task mixture of general and robotic data covering 2D detection and segmentation, hierarchical task decomposition, 3D object detection, spatial reasoning and trajectory prediction. Sizes are not published. Stage 2 (VLA): approximately 340 hours of high quality real robot data on the Unitree G1, plus systematically cleaned open source robot datasets. 12 Unitree G1 task datasets are published on Hugging Face.
Embodiments
Unitree G1 humanoid (12 task categories, single policy), Franka Panda in LIBERO simulation

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Published figures. These come from the model authors, not from this site.
Checkpoint download18.98 GB

No minimum GPU is stated. The repository documents CUDA 12.4, Python 3.10.18 and FlashAttention 2.5.6, and instructs users to set num_processes to the available GPU count for distributed training, without naming a device or a VRAM floor.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • LIBERO Object
    100%
    success rateSelf-reported.source
  • LIBERO Goal
    99.4%
    success rateSelf-reported.source
  • LIBERO Spatial
    99%
    success rateSelf-reported.source
  • LIBERO Average over 4 suites
    98.7%
    success rateSelf-reported by Unitree on the project page. Comparison table lists EO1 98.2, X-VLA 98.1, OpenVLA-OFT 97.1.source
  • LIBERO Long
    96.2%
    success rateSelf-reported. Weakest of the four suites, consistent with the general pattern that long horizon tasks are hardest.source

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Unitree reports a single end-to-end policy covering 12 categories of complex manipulation tasks on the G1 humanoid, including drawer opening and closing, connector plugging and unplugging, and pick and place. No numeric real robot success rates are published.

Fine tuning it yourself

Datasets must be in Hugging Face LeRobot v2.1 format, then converted with convert_lerobot_to_hdf5.py to HDF5 and finally built into RLDS via tfds build. Before training you set NUM_ACTIONS_CHUNK, ACTION_DIM, PROPRIO_DIM and ACTION_PROPRIO_NORMALIZATION_TYPE in constants.py. The shipped presets are NUM_ACTIONS_CHUNK 25 with ACTION_DIM 16 for G1 joint control, 25 with ACTION_DIM 23 for G1 EE-6D, and 8 with ACTION_DIM 7 for LIBERO. Fine-tuning starts from UnifoLM-VLM-Base weights. No demonstration count guidance is published.

Where it helps, where it does not

Strengths

  • Open weights plus training and inference code, with 12 matching Unitree G1 datasets on Hugging Face
  • Highest reported LIBERO average (98.7) in the vendor's own comparison table
  • Backed by the hardware vendor, so the G1 data, the datasets and the policy come from one source
  • Reference implementation reuses established stacks (Qwen2.5-VL, Isaac-GR00T, openvla-oft, InternVLA-M1)

Limits

  • No technical report and no arXiv paper, so architecture and training details exist only as code and a project page
  • No published parameter count for the VLA model, and the checkpoint format prevents deriving one
  • Spatial understanding results against Qwen2.5-VL-7B and Gemini Robotics-ER 1.5 are shown only as a figure image, with no numeric table anywhere
  • No real robot success rates published for the 12 G1 task categories
  • CC BY-NC-SA 4.0 forbids commercial use and requires share-alike
  • LIBERO comparison table is self-reported by Unitree, with no independent evaluation
  • Action chunk is inconsistent across artefacts: the shipped config.yaml sets action_horizon 16 while repository constants use 25 for G1 and 8 for LIBERO

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.

Download the clientDrive a real arm now