UnifoLM-WLA-1.0

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.

🔥 News

• Sep 20, 2026: 🚀 we released the model modules and training action expert code
• Sep 11, 2026: 🚀 we released the model weights of UnifoLM-ER-1
• Sep 11, 2026: 🚀 we released the model weights of UnifoLM-ER-Flow

📑 Open-Source Plan

• Code
  • Code for training action experts based on UnifoLM-ER models
  • Code for fine-tuning based on UnifoLM-WLA-Base
  • Code for LoRA fine-tuning
• Models
  • UnifoLM-ER-1
  • UnifoLM-ER-Flow
  • UnifoLM-WLA-Base
• Datasets
  • UniBot-V1 Challenge Dataset
  • UnifoLM-WBT-Dataset
  • UnifoLM-Dex1-Dataset

📘 Technical Documentation

• Robot Action, State, and Statistics Processing Specification
• Train an Action Expert from Scratch

Citation

@misc{unifolm-wla-1.0,
  author = {Unitree},
  title  = {UnifoLM-WLA-1.0: One Model Driven, Whole-Body Coordination},
  year   = {2026},
}

Acknowledgements

This project is built upon and continues the work of starVLA and Qwen-Image. We sincerely thank their authors and contributors for making their work publicly available.

License

Except where otherwise noted, this project is released under the Apache License 2.0. Third-party components remain subject to their original licenses. See NOTICE and THIRD_PARTY_LICENSES.md for attribution and details.