General-Purpose Humanoid Robot Foundation Model

UnifoLM-WLA-1.0One Model Drives All, Whole-Body Coordination

CodeComing soon ModelsComing soon DatasetsComing soon

UnifoLM-WLA-1.0 is Unitree Robotics' comprehensively upgraded, next-generation general-purpose humanoid robot foundation model with 6B parameters. Built on large-scale general multimodal perception and understanding data and interaction-centric world modeling, it substantially advances spatial perception and understanding, achieving leading results across multiple embodied reasoning benchmarks. Trained on approximately 2,500 hours of high-quality real-robot data, a single model coordinates 64 tasks spanning desktop manipulation and whole-body manipulation. It supports two-finger grippers and multiple five-finger dexterous hands, with strong generalization across tasks and end effectors.

6BVLA parameters

5M+ER samples

≈2,500hRobot data

ER Model · Embodied Reasoner

Understand the World,
Connect with Action.

A unified multimodal model brings together embodied reasoning, future dynamic-region prediction, and discrete action learning to jointly strengthen spatial perception, interaction prediction, and action generation, providing a unified vision–language–action representation for subsequent WLA training.

Embodied Reasoning

Across 16 multimodal perception and understanding benchmarks, UnifoLM-ER-1 leads open-source models on seven and delivers overall performance comparable to leading proprietary models. Built on Qwen3-VL-4B, UnifoLM-ER-1 is 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. These data are co-trained with general image–text data, preserving broad vision-language capabilities while substantially improving spatial understanding and reasoning in embodied environments.

Benchmark results

Model	Open Source	Spatial Understanding													Multimodal Understanding
Model	Open Source	RoboVQA	Ego-Plan2	RefSpatial- Bench	Where2Place	Pixmo-Point	BLINK	CV-Bench	EmbSpatial	RoboSpatial	SAT	VSI-Bench	VSR	ERQA	RealWorld QA	MME	MMMU_VAL
UnifoLM-ER-1-4B	Yes	62.4	55.1	61.7	82.0	73.8	93.4^†	88.6	88.9	73.1	76.0	54.2	88.1	50.0	69.8	2223.3	54.7
RoboBrain2.0-7B^*	Yes	30.0	33.23	42.2	63.6	54.7	83.9	85.7	76.3	54.2	75.3	36.1	84.0	/	69.4	2057.4	44.4
Robix-7B^*	No	63.6	/	/	41.9	29.5	87.6	86.5	77.4	/	71.1	44.6	83.3	42.5	70.7	2332.8	/
Pelican-7B^*	Yes	31.8	33.7	22.3	57.3	20.4	/	79.4	73.2	57.5	52.0	52.8	82.2	39.8	69.3	2141.9	51.1
Cosmos-R1-7B^*	Yes	38.8	26.0	5.6	2.9	8.2	/	76.7	68.9	42.4	82.7	25.4	82.4	/	67.6	2157.4	37.4
Cosmos3-Super-64B^*	Yes	/	/	57.0	71.0	/	90.3^†	88.0	/	70.0	/	60.9	/	51.2	/	/	/
Qwen3-VL-4B^*	Yes	47.7	40.7	46.6	63.0	48.3	85.0^†	85.1	79.6	61.7	68.7	59.3	81.6	41.3	71.0	2325.2	57.8
Qwen3-VL-8B^*	Yes	43.3	49.7	54.2	61.9	51.0	73.8^†	86.2	78.5	66.9	67.3	59.4	83.2	45.8	70.6	2412.5	62.3
Embodied-R1-3B^*	Yes	51.8	26.5	39.7	69.5	49.4	78.5^†	82.7	67.4	47.4	76.3	26.6	/	35.2	/	/	/
Embodied-R1.5-8B^*	Yes	61.0	53.8	54.2	74.0	64.8	83.0^†	86.9	78.1	69.7	74.7	56.1	/	46.0	/	/	/
Molmo2-ER-4B^*	Yes	/	/	52.5	54.0	/	85.7^†	87.8	78.8	/	78.0	74.5	/	46.8	/	/	/
Hy-Embodied-VLM-1.0 30B-A3B^*	Yes	/	49.6	53.4	65.0	64.6	87.3^†	89.7	82.7	69.4	78.0	/	/	60.8	/	/	/
MiMo-Emb-7B^*	Yes	62.0	43.0	48.0	63.6	42.35	81.3	88.2	76.2	61.7	78.6	48.5	79.0	46.7	66.3	2320.8	26.4
Thinker-4B^*	Yes	62.7	63.7	61.0	72.0	57.4	84.6	86.3	80.2	70.8	72.7	65.4	81.5	/	71.9	2323.4	46.2
Wall-OSS-0.5-3B^*	Yes	/	/	/	15.0	/	/	/	/	/	/	/	/	33	44	/	/
Lumo-1-Stage1-7B^*	No	/	/	51.0	69.1	/	82.4	86.4	75.6	62.6	74.7	/	/	/	/	/	/
Gemini-ER 2^‡	No	/	/	35.4	/	/	90.6^†	90.4	81.4	51.1	/	/	/	71.0	/	/	/
Gemini-ER 1.5^‡	No	/	/	41.8	48.0	/	/	83.6	73.4	57.7	62.0	39.9	/	47.0	/	/	/
Gemini 2.5 Pro^‡	No	/	/	33.6	37.0	/	88.6^†	85.9	78.0	71.3	74.7	51.1	/	56.0	/	/	/
Gemini 2.5 Flash^‡	No	/	/	41.2	48.0	/	80.3^†	85.5	76.2	73.4	73.3	45.3	/	47.5	/	/	/
Gemini 3.1 Pro^‡	No	/	/	70.0	61.0	/	86.1^†	88.6	/	65.1	/	47.5	/	65.2	/	/	/
GPT-5.6-sol^‡	No	/	/	58.3	51.1	/	85.6^†	85.2	80.7	66.8	21.3	/	/	64.8	/	/	/
GPT-6-Astra^‡	No	/	/	79.6	69.0	/	90.4^†	87.3	83.3	73.4	31.3	/	/	77.7	/	/	/

Results are sourced from the models' official technical reports or publicly available papers.

Results were obtained through tests using the models' official APIs.

BLINK scores are averaged over the Relative Depth and Spatial Relation subtasks only; all reported results were obtained in our own testing.

Dynamic Region Prediction

We use optical flow to extract dynamic regions that capture future scene changes, then train a VQ-VAE to encode them into fixed-length sequences of discrete tokens. Conditioned on the current image and a task description or action, the VLM directly predicts mask tokens for future dynamic regions, focusing on interaction subjects and the scene changes they induce to enable interaction-centric world modeling.

LIVE PIPELINE

01 / OPTICAL-FLOW SUPERVISION

IMAGE_t0

IMAGE_t1

ESTIMATEOptical
Flow

Optical flow from t0 to t1 — OPTICAL FLOW

Dynamic-region mask overlay — DYNAMIC MASK

DISCRETIZEVQ-VAE

TaskORAction

IMAGE_t0 + CONDITIONVLM

02 / FUTURE CHANGE PREDICTION

08213704 19421126 FUTURE MASK TOKENS

Clean table01

Fold towel02

Install phone03

Place plates04

Discrete Action Learning

We partition the unified action space into three components: end-effector (EEF) poses, end-effector joints, and lower-body joints. A separate residual vector quantization (RVQ) model is trained for each component to discretize action sequences. Building on UnifoLM-ER-1, we introduce discrete action tokens and mask tokens for future dynamic regions, jointly aligning visual, language, and action representations within a single VLM to obtain the UnifoLM-ER-Flow model.

LIVE ENCODING

EEF trajectory

RVQ

<EEF_START>1842073156<EEF_END>

Gripper / dexterous hand

RVQ

<HAND_START>0429511638<HAND_END>

Lower-body motion

RVQ

<LOWER_START>2711440322<LOWER_END>

Temporal alignmentShared timesteps · Synchronously fed into VLM

WLA Model

Multi-Source Data,
One Model.

UnifoLM-WLA-1.0 builds on the UnifoLM-ER-Flow multimodal backbone and incorporates an MMDiT action expert. It is trained on approximately 2,500 hours of high-quality real-robot data—including the Unitree Open Datasets, BitRobot-HIW-500—covering diverse robot embodiments and operational scenarios. Through a unified action space, it enables cross-embodiment prior transfer and jointly models perception and understanding, interaction prediction, and action generation, balancing embodied manipulation capabilities with general multimodal perception and reasoning.

Training Setup

LIVE GRAPH

Language prediction

Discrete actions

Continuous actions

UnifoLM-ER-Flow

Action ExpertMMDiT flow decoder

Image encoder

Prompt

State

Noise Actions

Real-robot tests · Unitree G1

One Model Driven,
Multiple Tasks, Multiple End-effectors.

A single UnifoLM-WLA-1.0 model supports both tabletop and whole-body manipulation, demonstrating smooth task execution across diverse real-robot evaluations. The videos below showcase real-robot evaluation results.

10 Whole-Body Manipulation Tasks

Scroll to explore

54 Table-top Manipulation Tasks