WholebodyVLA README

The project's README from github.com/OpenDriveLab/WholebodyVLA. As one markdown file: README.md.

WholeBodyVLA Logo

Towards Unified Latent VLA for Whole-body Loco-manipulation Control

Paper Home

✒️ Haoran Jiang*, Jin Chen*, Qingwen Bu, Li Chen, Modi Shi, Yanjie Zhang, Delong Li, Chuanzhe Suo, Chuang Wang, Zhihui Peng†, Hongyang Li†

📧 Primary Contact: Haoran Jiang ([email protected]).

🔥 Highlights

📋 Overview

WholeBodyVLA is a unified Vision-Language-Action framework for large-space humanoid loco-manipulation. It learns unified latent actions from action-free egocentric videos through a Latent Action Model (LAM), and employs a loco-manipulation-oriented (LMO) RL policy for precise and stable whole-body coordination. The system encodes egocentric images and language instructions into latent action tokens, which are decoded into dual-arm joint actions and locomotion commands, enabling end-to-end control for complex loco-manipulation tasks.

WholeBodyVLA Method

See more on project website.

📝 Note: We currently have no concrete timeline for open-sourcing the codebase. This repository now serves as a collection of resources and references for the whole-body humanoid VLA research community. We welcome discussion and collaboration!

Let's go for VLA on humanoids!

Awesome Vision–Language–Action for Humanoid Robots

A curated list of research on Vision–Language–Action (VLA) models and related for humanoid robots, with a focus on loco-manipulation task.

Continuously updating...

Perception & Planning for Humanoids

Vision-Based Perception for Planning

Manipulation

Other Tasks

MoCap-Based Planning

Generative Motion and Trajectory Planning

Whole-Body Controller for Loco-Manipulation

Behavior Foundation Models / Universal Whole-Body Tracking

Upper-Body Centric

Hardware Design

Data Collection Systems

Capability Extension

Generalist Vision–Language–Action Models

Manipulation

Learning from Human Videos

Navigation

References

We refer to and recommend several curated paper lists and repositories:

Contributors

This project is contributed by: Jin Chen, Yucheng Huang, Haoran Jiang, Yixuan Pan, Shijia Peng, Jialong Zeng, Hai Zhang.

All names are listed in alphabetical order by last name.

Citation

If you find this repository helpful, please consider citing:

@article{jiang2025wholebodyvla,
  title={WholeBodyVLA: Towards Unified Latent VLA for Whole-Body Loco-Manipulation Control}, 
  author={Jiang, Haoran and Chen, Jin and Bu, Qingwen and Chen, Li and Shi, Modi and Zhang, Yanjie and Li, Delong and Suo, Chuanzhe and Wang, Chuang and Peng, Zhihui and Li, Hongyang},
  journal={arXiv preprint arXiv:2512.11047},
  year={2025}
}

@article{chen2025intelligent,
  title={Intelligent Robot Manipulation Requires Self-Directed Learning},
  author={Chen, Li and Sima, Chonghao and Chitta, Kashyap and Loquercio, Antonio and Luo, Ping and Ma, Yi and Li, Hongyang}, 
  year={2025}
}

WholebodyVLA (OpenDriveLab/WholebodyVLA)

Searching Database…

580 ·GitHub
WholebodyVLA
What is this?WholeBodyVLA is an ICLR 2026 research project introducing a unified Vision-Language-Action framework for closed-loop humanoid whole-body loco-manipulation control, and this repository doubles as a curated, continuously updated list of VLA papers for humanoid robots.
Is this a working codebase or just a paper and reading list?It's mainly a paper overview plus an 'Awesome Vision-Language-Action for Humanoid Robots' reading list; the WholeBodyVLA code itself has not been released.
How do I actually use this repository?There's nothing to install — you browse the categorized paper list (perception/planning, whole-body controllers, generative motion planning, MoCap-based planning) and follow each entry's arXiv, Code, or Demo link.
Can I run the WholeBodyVLA model myself right now?No — the authors give no release timeline for the codebase, so only the paper's description and the linked reference list are currently available.
How does WholeBodyVLA technically work?It learns unified latent actions from action-free egocentric videos via a Latent Action Model (LAM), encoding images and language instructions into latent action tokens that decode into dual-arm joint actions and locomotion commands, guided by a loco-manipulation-oriented (LMO) RL policy for stable whole-body coordination.
How is the awesome list organized?Papers are grouped by topic such as vision-based perception for manipulation, MoCap-based planning, generative motion/trajectory planning, and whole-body controllers, each tagged with venue, year, and links to arXiv, code, or demos when available.
Is the list still being updated?Yes — it's explicitly described as 'Continuously updating' and already includes entries from as late as early 2026.
What are similar projects?awesome-rsi, Motus2, RoboICL, Zero-WAM, lerobot, transferlearning, Embodied-AI-Start, Embodied-AI-Guide
Which topics is it in?vision-language-action, robot, embodied-ai, papers, paper, robotics, academic, reinforcement-learning, multimodal, awesome, artificial-intelligence, computer-vision
Creator
OpenDriveLab
Size
44.9 MB
Stars
580
Updated
CommitMay 25
Age
10mo
Commits · 90 Days