Unified Vision-Language-Action Model

OneVLA: A Unified Framework for Embodied Tasks

A single VLA architecture for both navigation and manipulation, without task-specific model variants.

Lingfeng Zhang1,2,3, Xiaoshuai Hao3,†,‡, Yingbo Tang3,4, Lei Zhou3, Shuyi Zhang4, Jinkun Liu1, Hongsheng Li1, Chenhao Zhang5, Qiang Zhang6, Hangjun Ye3, Xiaojun Liang2, Long Chen3, Wenbo Ding1,†
1Tsinghua University 2Pengcheng Laboratory 3Xiaomi EV 4Institute of Automation, Chinese Academy of Sciences 5Peking University 6HKUST(GZ)
†Corresponding authors ‡Project leader
OneVLA overview and embodied tasks
Navigation

Follow language instructions through unseen environments.

Manipulation

Execute continuous 7-DOF robot control from visual observations.

One model, two embodied capabilities

OneVLA unifies navigation and manipulation with a shared vision-language backbone, a flow-matching action head, and an 11-dimensional action space that covers both discrete navigation commands and continuous manipulation actions.

Abstract

Navigation and manipulation are fundamental capabilities of embodied intelligence, enabling robots to interpret natural language commands and interact physically with their surroundings. However, current Vision-Language-Action models often rely on task-specific architectures, specializing in either navigation or manipulation and limiting their use as general-purpose robotic agents.

OneVLA introduces a unified architecture that integrates these distinct tasks into a single framework. It uses a unified action head to generate both navigation and manipulation actions without task-specific variants, and adopts a multi-stage progressive training strategy with curated data construction and Chain-of-Thought fine-tuning. Experiments in simulated and real-world environments show that OneVLA achieves state-of-the-art performance and that navigation and manipulation can mutually reinforce each other during joint training.

Method Overview

OneVLA learns shared vision-language representations and predicts task-specific actions through a single unified action head.

Overall framework of OneVLA
Overall framework of OneVLA. Multimodal observations and language instructions are encoded by a unified vision-language backbone, then decoded into navigation or manipulation actions in an 11-dimensional action space.
01

Unified Inputs

Multi-view RGB observations, language instructions, and optional robot state are processed by one shared backbone.

02

Dual Output

The model generates language reasoning and future robot actions in a single forward pass.

03

Unified Action Space

Four navigation dimensions and seven manipulation dimensions are concatenated into one 11D action representation.

04

Task-Aware Loss

Dimension-specific masks supervise only the relevant action dimensions for each embodied task.

Progressive Training

A three-stage curriculum first builds manipulation skill, then introduces navigation for cross-task transfer, and finally refines the model with Chain-of-Thought data.

Multi-stage training strategy of OneVLA

Results

OneVLA achieves strong performance across navigation and manipulation while using a single 3B unified model.

68.6 VLN-CE R2R

+21.5 over UniVLA-Navi with task-specific fine-tuned heads.

58.2 VLN-CE RxR

+31.9 over UniVLA-Navi on unseen continuous navigation.

64.5 SimplerEnv Avg.

+29.1 over UniVLA-Mani across representative manipulation tasks.

3B Unified Model

Outperforms 7B single-task and cross-task baselines without task-specific variants.

Qualitative Visualization

Representative examples across simulator and real-world navigation and manipulation.

Simulator navigation visualization
Simulator Navigation
Simulator manipulation visualization
Simulator Manipulation
Real-world navigation visualization
Real-World Navigation
Real-world manipulation visualization
Real-World Manipulation

BibTeX

@article{zhang2026onevla,
  title={OneVLA: A Unified Framework for Embodied Tasks},
  author={Zhang, Lingfeng and Hao, Xiaoshuai and Tang, Yingbo and Zhou, Lei and Zhang, Shuyi and Liu, Jinkun and Li, Hongsheng and Zhang, Chenhao and Zhang, Qiang and Ye, Hangjun and Liang, Xiaojun and Chen, Long and Ding, Wenbo},
  journal={Preprint},
  year={2026}
}