UnifoLM-WLA-1.0 – Unitree Robotics' Embodied Multimodal Large Model

1. What is UnifoLM-WLA-1.0

UnifoLM-WLA-1.0 is a 6B-parameter embodied multimodal large model launched by Unitree Robotics, trained on approximately 2500 hours of real-robot operation data. It unifies visual perception, language understanding, and action execution into a single modeling framework, enabling a single model to handle 64 desktop and full-body mobility operation tasks. The model employs a unified VLA architecture, using a vision-language model as its backbone. It converts continuous control problems into token prediction problems through RVQ discrete action representation and introduces a future dynamic prediction mechanism in the ER-Flow layer, allowing robots to "imagine" the scene changes caused by an action before executing it. The model has open-sourced approximately 4B parameters of the UnifoLM-ER-1 perception-inference model and ER-Flow action modeling weights. The complete system and post-training code will be gradually released, providing a fully open-source, locally deployable foundation solution for the field of embodied intelligence.

Image source: Official article
Image source: official article

Technical positioning and domain: It belongs to the intersection of embodied intelligence and multimodal large models, focusing on the unified modeling of visual-language-action for robotic operation tasks. It is one of the few fully open-source end-to-end VLA systems in this field trained on large-scale real-robot data.

Development background: Developed independently by Unitree Robotics, a company with deep expertise in humanoid robot hardware, known for launching several well-known humanoid robot products such as the G1 and H1. Building on this foundation, Unitree has extended its technology stack into embodied intelligence algorithms, releasing the weights of the UnifoLM series of models.

Core value: It addresses the issue of error accumulation caused by the modular separation of perception, planning, and control in traditional robotic operations. By achieving end-to-end training with a single model, it enables unified execution across multiple tasks. Its fully open-source strategy lowers the barrier to embodied intelligence research, allowing universities, laboratories, and small and medium-sized enterprises to conduct VLA algorithm research based on real robot data.

Technical features: It uses a unified VLA architecture to end-to-end train a 6B-parameter model, encoding dexterous hand movements, end-effector poses, and lower-limb joints into action tokens through RVQ discretization. The ER-Flow layer predicts both actions and future dynamic regions simultaneously during inference using a masked token mechanism, achieving a closed-loop control system that "predicts first, then acts."

2. Key Features

• Embodied Perception and Reasoning: The underlying ER-1 model is based on Qwen3-VL-4B and has been reinforced with over 5 million samples. It can identify object locations, spatial relationships, and manipulable areas, answering spatial cognition questions such as "Where is the target?" and "What is the relationship between objects?" to provide a perceptual foundation for operational decisions. It has achieved leading performance in seven open-source model evaluations, including RefSpatial-Bench.

• Unified Execution of Vision-Language-Action: Directly maps task instructions to a sequence of actions executable by the robot. A single model manages 54 desktop tasks and 10 full-body mobility operations, eliminating the need for separately training strategy models for each task and significantly reducing the engineering complexity of multi-task deployment.

• Interaction Outcome Prediction: Through the future dynamic prediction mechanism in the ER-Flow layer, predicts scene changes caused by the operation before executing the action, enabling the robot to "imagine first, then act," reducing ineffective attempts and collision risks, and improving operational success rates.

• Full-Body Coordinated Operation: Integrates upper-limb fine manipulation and lower-limb mobility control for overall coordination, enabling the robot to complete complex tasks such as making the bed, folding clothes, and taking out trash that require close coordination between movement and manipulation. This breaks through the limitations of traditional robotic arms fixed to a stationary base.

• Unified Modeling of Discrete Actions: Utilizes RVQ (Residual Vector Quantization) to discretize end-effector pose, dexterous hand joints, and lower-limb joints into action tokens, transforming continuous control problems into token prediction problems that language models excel at. This allows for unified processing of visual, textual, and action information within the same sequence.

• Cross-Body Deployment Capability: The model is deeply optimized for Unitree G1/H1 humanoid robots, with decoded action instructions capable of being integrated into the robot's motion control interface. It also supports verification in simulation environments first, providing a flexible deployment path for migration across different hardware platforms.

3. How to Use

1. Get the Code and Install Dependencies: Visit the project homepage (unigen-x.github.io/unifolm-wla.github.io) or the GitHub repository (github.com/unitreerobotics/unifolm-wla), clone the code repository according to the README instructions, and install the required dependencies. It is recommended to use Python 3.10+ and configure a PyTorch environment that supports CUDA.

2. Download Model Weights: Pull the open-source model weights for UnifoLM-ER-1 and UnifoLM-ER-Flow from HuggingFace using the huggingface-cli download command. The two models combined have approximately 4B parameters, so ensure that there is sufficient local storage space.

3. Run Perception Inference Examples: First, use the provided inference scripts in the project to load the ER-1 model. Input the robot's current visual feed and task description to test whether perception and inference outputs such as object localization and spatial relationship understanding are functioning correctly. This step is used to verify the correctness of the environment configuration.

4. Integrate with Action Prediction Module: Load the ER-Flow model and input the current image, task instruction, and candidate actions. The model will output discrete action tokens and predictions for future dynamic regions. You must use the official RVQ decoding method to convert the tokens back into continuous control signals.

5. Adapt to Real Robot Control Interfaces: Connect the decoded action instructions to the motion control interface of Unitree G1/H1 robots. Perform small-scale verification in a simulation environment or on the actual robot before deployment. Note that the model weights themselves do not include a complete robot control system, so the relevant control logic must be implemented separately.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: For research teams and university labs that require full control over code and data, UnifoLM-WLA-1.0, with its Apache 2.0 license and hierarchical model design, serves as an ideal foundation for VLA research, especially for teams using Unitree robots as their hardware platform. While π0.5 excels in open-world generalization, its closed ecosystem makes it more suitable for commercial teams that use it through partner collaborations. For lightweight desktop manipulation tasks with no hardware platform restrictions, OpenVLA's cross-ontology pre-training advantages are worth considering. For research institutions with ample computational resources and seeking the benefits of large-scale pre-training, the RT-2 paper offers valuable reference.

6. Editor's Summary

UnifoLM-WLA-1.0 demonstrates a differentiated technical approach in the field of embodied multimodal large models. From the perspective of technological innovation, the future dynamic prediction mechanism in the ER-Flow layer is a key design that distinguishes it from most VLA models — by predicting both actions and dynamic regions simultaneously through masked tokens, the model completes the imagination and reasoning of scene changes before executing actions. This "predict first, act later" closed-loop architecture is relatively rare in the open-source domain. The layered design (ER-1 perception foundation + ER-Flow action layer) also reflects practical engineering considerations, allowing developers to flexibly combine and use capabilities from different layers as needed. The training strategy using approximately 2,500 hours of real robot data ensures that the action outputs are closer to the actual dynamics of real robots, which is an advantage over approaches that heavily rely on simulated data.

In terms of practical value, the fully open-source (Apache 2.0) strategy significantly reduces the research threshold for embodied intelligence, enabling academic laboratories, small and medium-sized teams, and individual developers to access the complete technical stack of end-to-end VLA systems. The model's ability to handle 64 tasks under a single framework also saves substantial training costs in practical deployment scenarios. In terms of hardware compatibility, its deep integration with Unitree's G1/H1 creates a synergistic ecological effect, allowing users to directly validate algorithm performance on real humanoid robots.

This model is primarily suitable for four groups of people: academic laboratories focusing on embodied intelligence as a core research direction, commercial teams developing applications based on Unitree robots, technical enthusiasts seeking a reference implementation for end-to-end VLA systems, and algorithm engineers interested in starting with modular perception and reasoning. As the full system and post-training code are gradually made available, UnifoLM-WLA-1.0 is expected to become an important foundational infrastructure for the domestic open-source ecosystem in embodied intelligence. The evolution of subsequent versions in terms of task coverage and cross-body generalization capabilities is worth watching.

7. Application Scenarios

• Home Service Scenario: Performing daily chores such as folding clothes, making the bed, and taking out the trash was the key demonstration scenario at the launch event. Users can directly instruct the Unitree robot equipped with UnifoLM-WLA-1.0 through natural language commands. The robot relies on its perception and reasoning capabilities to locate objects, plan actions, and avoid collisions using future dynamic prediction. Full-body coordinated control makes these complex tasks that require movement and operation coordination a reality.

• Industrial Manufacturing Scenario: Executing repetitive operational tasks such as part picking, tool delivery, and simple assembly on the production line. Manufacturing companies can deploy the model locally to protect production line data privacy, while leveraging its single-model multi-task capability to quickly switch between different workstation tasks, without needing to train separate strategies for each operation, thereby reducing the cost of production line automation upgrades.

• Logistics and Warehouse Scenario: Performing warehouse operations that require coordinated movement and manipulation, such as cargo sorting, package handling, and shelf organizing. The model's spatial perception and reasoning capabilities enable it to identify shelf compartments and cargo locations, while full-body coordinated manipulation allows the robot to move through aisles and complete pick-and-place actions, increasing the level of automation in warehouse operations.

• Research and Education Scenario: As a fully open-source end-to-end VLA base, it provides universities and laboratories with a platform to research embodied intelligence algorithms and train embodied reasoning data. Researchers can study perception and reasoning capabilities independently using ER-1, or combine it with ER-Flow to explore action prediction mechanisms, with its layered design offering flexible entry points for algorithm research.

• Commercial Guidance and Service Scenario: Performing service-oriented tasks involving human interaction, such as guiding visitors and delivering items, in exhibition halls, shopping malls, and other locations. The model's language understanding capabilities support the interpretation of visitor instructions, while its embodied manipulation capabilities enable physical interactions such as item delivery, offering commercial service robots a more natural user experience.

8. FAQ

Q: What is the core difference between UnifoLM-WLA-1.0 and regular VLA models?
A: The core difference lies in the future dynamic prediction mechanism of the ER-Flow layer. During inference, the model not only predicts action tokens but also simultaneously predicts future dynamic region tokens, forming a closed-loop of "first imagining the consequences of an action, then executing it." This design enables the model to anticipate scene changes caused by actions, reducing unnecessary attempts and collision risks, a capability that most open-source VLA models lack.

Q: What is included in the open-source portion, and when will the full system be released?
A: Currently, the open-source portion includes the weights of the UnifoLM-ER-1 perception-inference model and the ER-Flow action modeling model, along with corresponding inference scripts and RVQ decoding solutions. The full 6B system and post-training code will be released gradually, with the exact timeline to be announced officially.

Q: Can the model be adapted to robot hardware from non-Unitree brands?
A: The model's output action tokens are deeply optimized for Unitree robot control interfaces. To migrate to other brands' hardware, users must implement their own RVQ decoding and control signal conversion logic. It is recommended to first verify compatibility in a simulation environment before deploying on specific hardware platforms.

Q: Is the training data based on simulated data or real robot data?
A: The model is trained on approximately 2500 hours of real robot operation data, rather than purely simulated data. This makes the action outputs more aligned with the actual dynamics of real robots, enhancing physical feasibility. However, the training data mainly comes from Unitree's own collection, resulting in relatively limited cross-platform generalization capabilities.

Q: What hardware configuration is required to deploy UnifoLM-WLA-1.0?
A: The two open-source models combined have approximately 4B parameters, and it is recommended to use a GPU with at least 24GB of VRAM for inference. The hardware requirements for the full 6B system will be confirmed after the official release. The official recommendation is to use a CUDA-supported environment with Python version 3.10 or higher.

Q: Must ER-1 and ER-Flow be used together, or can they be used separately?
A: They can be used independently. ER-1 focuses on perception and reasoning and can be used alone for tasks such as target localization and spatial relationship understanding. ER-Flow relies on the perception output from ER-1 to perform action modeling predictions. Users can flexibly choose which level to use based on their specific needs.

9. Project Links

• Project Website: https://unigen-x.github.io/unifolm-wla.github.io/
• GitHub Repository: https://github.com/unitreerobotics/unifolm-wla
  • https://huggingface.co/unitreerobotics/UnifoLM-ER-Flow
  • https://huggingface.co/unitreerobotics/UnifoLM-ER-1