Unitree Releases UnifoLM-WLA-1.0, a 6B Model for Humanoids

Unitree Robotics has released the UnifoLM-WLA-1.0 model for performing manipulation and locomotion tasks on humanoid robots. This model is 6B parameters in scale and adopts a new architecture that, unlike existing video generation methods, compresses and predicts only the dynamic region of future scenes as discrete tokens.

Core Technology and Architecture

• Dynamic Region Prediction: Instead of generating full videos, it tokenizes Optical Flow-based dynamic regions using VQ-VAE for prediction, thereby reducing computational costs and focusing on information necessary for robot policy learning.
• Three-Stage Structure: Based on the embodied cognition VLM UnifoLM-ER-1 (4B), it goes through UnifoLM-ER-Flow for dynamic/action token alignment, culminating in the final policy model UnifoLM-WLA-1.0 (6B).
• Action Tokenization: Continuous trajectories are discretized into tokens using RVQ (Residual Vector Quantization), with the position and rotation of the end-effector learned separately using 256 codes each.

Performance and Data

• Training Data: A total of 2,500 hours of real-world data was used.
• Real-World Demos: Videos were released showing the Unitree G1 robot performing a total of 64 tasks, including 54 tabletop manipulation tasks and 10 whole-body locomotion manipulation tasks.

Release Scope and License

• Released Items: Model weights for UnifoLM-ER-1 and UnifoLM-ER-Flow, along with some datasets (UniBot-V1, WBT), have been released under the Apache 2.0 license.
• Unreleased Items: The weights for the final model UnifoLM-WLA-Base (6B), post-training code, and the full training dataset have not yet been released.
• License Note: While model weights are Apache 2.0, GitHub repository documentation and some datasets are under CC BY-NC-SA 4.0 or have no license specified, so individual verification is required for commercial use.