过去几年间，多模态大语言模型在理解图像、视频和真实世界场景方面能力日益增强。它们能够识别物体、推理空间关系、回答视觉问题，并解决复杂的多模态推理任务。
但对于具身智能而言，理解世界仅仅是第一步。一个真正的具身智能体还需要理解任务目标、在物理世界中采取行动，并在不同的机器人本体、环境和任务之间实现泛化。
这正是 Qwen-VLA 背后的研发动机。
Qwen-VLA 是一个通用型的视觉-语言-动作模型。它基于 Qwen 多模态骨干网络构建，将视觉感知、语言理解和空间推理扩展至连续动作生成与轨迹预测。换言之，该模型不仅能看、能思考，还能开始行动。
一个模型应对多种具身任务
传统的具身智能系统往往高度专用化：一个模型负责桌面操作，另一个负责导航，再一个专门针对特定机器人平台。这种方法在单个任务上可能表现良好，但难以扩展到更广泛的任务、多样化的环境或不同的机器人本体。
Qwen-VLA 探索了一条更统一的路径：
能否用一个通用策略模型同时支持机器人操作、视觉-语言导航以及跨本体控制？
在 Qwen-VLA 中，机器人操作和视觉-语言导航被统一到同一框架下：给定视觉观测、语言指令和本体特定条件，模型预测下一步动作或轨迹。Qwen 多模态骨干网络负责理解视觉和语言输入，而动作解码器则生成连续动作。
训练：从语言先验到闭环控制
Qwen-VLA 的核心并非简单地在多模态模型上附加一个动作头。更重要的是，它构建了一个覆盖多样化任务、环境和机器人本体的联合训练系统。完整的训练流程分为四个阶段，从语言先验逐步过渡到闭环控制。
数据
预训练数据涵盖五大来源：
机器人操作轨迹构成基础，覆盖桌面、移动、双臂及灵巧操作。公开数据总计超过10,000小时，辅以1,000多小时的内部真实机器人轨迹和超过800万条合成仿真轨迹。
人类第一人称视角数据从开放世界环境中提供了更丰富的物体、场景和手部动作先验知识。我们整合了Ego4D、EPIC-KITCHENS、EgoDex（829小时）、EgoVerse（1,300多小时、1,965个任务、240个场景）以及Xperience。
合成仿真数据填补了长尾场景的空白。视觉条件数据涵盖20个桌面场景、200种配置、450个任务和359,848条成功轨迹。文本到动作数据覆盖6种模板×6种单臂机器人，产生约720万条轨迹和超过14,000小时的数据。
视觉语言导航数据提供了长程轨迹规划和指令跟随能力。
通用视觉语言数据保留了多模态理解、空间定位和指令跟随能力。我们还构建了约48,000条精细动作描述，涵盖13个维度的标注，将自然语言与具体的执行细节对齐。
四阶段训练#
核心思路：先学习从语言生成动作结构，再学习让这些动作适应视觉环境。
第一阶段：T2A（文本到动作预训练）。像“拿起红色杯子”这样的指令只有几个词，但对应的机器人动作却是一条高维连续轨迹。Qwen-VLA将此视为从语言到动作的一种解压缩过程。在T2A阶段，我们冻结VLM，仅使用语言和具身提示词（不含任何图像）来训练动作解码器。
第二阶段：CPT（持续预训练）。我们同时解冻VLM和动作解码器，并在完整的混合多模态数据上进行联合训练。此阶段将T2A阶段的语言-动作先验知识锚定到具体的视觉场景中，同时让主干网络适应具身感知，从而得到Qwen-VLA-Base。
第三阶段：SFT（监督微调）。从 CPT 检查点出发，我们分叉为两条路线：多任务 SFT 联合微调操作、导航、VQA 和空间定位；真实机器人 SFT 则基于内部遥操作数据进行微调，用于物理部署。
第四阶段：RL（强化学习）。从 SFT 检查点出发，我们使用 PPO 直接在模拟环境中优化闭环任务成功率，最终得到 Qwen-VLA-Instruct 模型。RL 仅在 SimplerEnv 中进行，但实验表明其收益可迁移至未见过的环境和机器人形态。
单一通用模型可媲美甚至超越专用模型#
实验结果表明了 Qwen-VLA 作为通用策略模型的潜力。单一模型可覆盖多个操作基准，包括 LIBERO、Simpler、RoboCasa 和 RoboTwin，并在多项任务上接近或超越专用策略模型。
| 基准 | 最佳专用模型 | Qwen-VLA |
|---|---|---|
| LIBERO | ABot-M0 98.6% | 97.9% |
| RoboCasa-GR1 | ABot-M0 58.3% | 56.7% |
| Simpler-WidowX | StarVLA-OFT 64.6% | 73.7% |
| RoboTwin-简单 / 困难 | ABot-M0 86.0% / 85.0% | 86.1% / 87.2% |
在机器人操作基准上，Qwen-VLA-Instruct 在 LIBERO 上达到 97.9%，在 Simpler-WidowX 上达到 73.7%，在 RoboTwin-简单 / 困难上分别达到 86.1% / 87.2%。许多对比方法是为单个基准微调的专用模型，而 Qwen-VLA 是在单一框架下训练的统一通用模型。
在视觉语言导航（VLN-CE）任务上，Qwen-VLA-Instruct 在 R2R Val-Unseen 上达到 69.0% 的 Oracle 成功率和 57.5% 的成功率，在更具挑战性的 RxR Val-Unseen 上达到 59.6% 的成功率和 47.8% 的 SPL，超越了所有开源基线模型。
在真实世界的 ALOHA 双臂实验中，Qwen-VLA 预训练模型在域内任务上平均成功率达到 83.6%，在域外任务上平均成功率达到 76.9%，大幅超越从零开始训练（48.5% / 36.2%）和 $\pi{0.5}$π 0.5（71.6% / 41.5%）的结果。
真实世界域外泛化能力#
我们同样关注 Qwen-VLA 在真实机器人上的泛化表现。
在真实世界的 ALOHA 双臂机器人实验中，Qwen-VLA 展现了对未见过的颜色、物体、背景、位置和语言指令的泛化能力。与从头训练的模型相比，使用 Qwen-VLA 预训练的模型在真实世界的分布外场景下表现出明显的性能提升。
这部分最好通过视频来展示。以下演示均使用 Qwen-VLA-Base 模型进行测试。当要求“拿起绿色球”或“拿起蓝色球”时，模型能够根据颜色特定的指令正确执行动作。当出现玩具、蔬菜或太阳镜等未见过的物体时，模型仍能遵循语言指令抓取或移动它们。当背景、光照和桌面布局发生变化时，模型保持相对稳定。对于“整理桌面”这类组合型任务，模型能够识别多个目标并执行多步骤操作。
与单纯的表格相比，这些视频更能体现 Qwen-VLA 的核心价值：
该模型并非仅仅在固定环境中记忆动作模板，而是在学习理解目标，并在真实世界的各种变化条件下执行动作。
动态场景中的零样本泛化
除了静态桌面操作之外，Qwen-VLA 在动态操作任务中也展现出了零样本泛化能力。
在 DOMINO 动态操作基准测试中，Qwen-VLA-Instruct 并未针对该基准进行专门微调，但仍取得了 26.6% 的成功率和 39.5 的操作得分，超越了多种标准 VLA 基线模型，甚至优于一些专门用于动态操作任务的专家模型。
这表明该模型不仅在静态场景中学习抓取模板，还从空间理解到运动控制层面，习得了更具迁移性的动作先验知识。基于视觉观察、语言目标及其动作生成能力，模型能够直接生成连贯的动作序列，并在动态交互窗口内完成任务。
从多模态理解到具身智能
Qwen-VLA 是通义千问（Qwen）多模态能力向具身智能领域的自然延伸。
过去，多模态模型主要侧重于理解世界。通过 Qwen-VLA，我们进一步探索模型如何基于视觉和语言在物理世界中生成动作。
Qwen-VLA 统一了机器人操控、视觉语言导航以及跨本体控制。它将 Qwen 的视觉理解与空间推理能力连接到连续动作生成。通过在真实机器人数据、人类第一人称数据、合成仿真数据以及通用视觉语言数据上进行联合预训练，它学习到了更通用的具身经验。同时，它也展示了通用策略模型在操控基准测试、真实世界分布外泛化以及零样本动态操控方面的潜力。
具身智能仍处于早期阶段。长程真实世界任务、故障恢复、持续学习以及更复杂的人-机-环境交互仍然充满挑战。但 Qwen-VLA 指明了清晰的下一步：
模型不仅应理解世界——它们还应学会在其中行动。
@article{qwenvla, title={Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments}, author={Qwen Team}, year={2026}, eprint={2605.30280}, archivePrefix={arXiv}, primaryClass={cs.RO}, url={https://arxiv.org/abs/2605.30280},}
Over the past few years, multimodal large language models have become increasingly capable of understanding images, videos, and real-world scenes. They can recognize objects, reason about spatial relationships, answer visual questions, and solve complex multimodal reasoning tasks.
But for embodied intelligence, understanding the world is only the first step. A truly embodied agent also needs to understand task goals, take actions in the physical world, and generalize across different robot embodiments, environments, and tasks.
This is the motivation behind Qwen-VLA.
Qwen-VLA is a general-purpose Vision-Language-Action model. Built upon the Qwen multimodal backbone, it extends visual perception, language understanding, and spatial reasoning into continuous action generation and trajectory prediction. In other words, it allows the model to not only see and think, but also begin to act.
One Model for Multiple Embodied Tasks#
Traditional embodied AI systems are often highly specialized: one model for tabletop manipulation, another for navigation, and yet another for a specific robot platform. This approach can work well for individual tasks, but it does not scale easily to broader tasks, diverse environments, or different robot embodiments.
Qwen-VLA explores a more unified direction:
Can a single generalist policy model support robotic manipulation, vision-language navigation, and cross-embodiment control at the same time?
In Qwen-VLA, robotic manipulation and vision-language navigation are formulated under the same framework: given visual observations, language instructions, and embodiment-specific conditions, the model predicts the next action or trajectory. The Qwen multimodal backbone understands the visual and language inputs, while an action decoder generates continuous actions.
Training: From Language Priors to Closed-Loop Control#
The core of Qwen-VLA is not simply attaching an action head to a multimodal model. More importantly, it builds a joint training system that covers diverse tasks, environments, and robot embodiments. The full training pipeline progresses through four stages, from language priors to closed-loop control.
Data#
The pretraining data spans five major sources:
Robot manipulation trajectories form the foundation, covering tabletop, mobile, dual-arm, and dexterous manipulation. The public data totals over 10,000 hours, supplemented by more than 1,000 hours of internal real-robot trajectories and over 8 million synthetic simulation trajectories.
Human egocentric data provides richer object, scene, and hand-action priors from open-world environments. We incorporate Ego4D, EPIC-KITCHENS, EgoDex (829 hours), EgoVerse (1,300+ hours, 1,965 tasks, 240 scenes), and Xperience.
Synthetic simulation data fills long-tail gaps. Vision-conditioned data covers 20 tabletop scenes, 200 configurations, 450 tasks, and 359,848 successful trajectories. Text-to-action data spans 6 templates × 6 single-arm robots, yielding about 7.2 million trajectories and over 14,000 hours.
Vision-language navigation data provides long-horizon trajectory planning and instruction-following capabilities.
General vision-language data preserves multimodal understanding, spatial grounding, and instruction following. We also build around 48,000 fine-grained action descriptions annotated across 13 dimensions, aligning natural language with concrete execution details.
Four-Stage Training#
The key idea: first learn to generate action structures from language, then learn to adapt those actions to the visual environment.
Stage I: T2A (Text-to-Action Pretraining). An instruction like “pick up the red cup” is just a few words, but the corresponding robot action is a high-dimensional continuous trajectory. Qwen-VLA treats this as a form of decompression from language to action. In T2A, we freeze the VLM and train only the action decoder on language and embodiment prompts without any images.
Stage II: CPT（Continual Pretraining）. We unfreeze both the VLM and action decoder and jointly train on the full multimodal data mixture. This stage grounds the language-action priors from T2A in concrete visual scenes while adapting the backbone to embodied perception, producing Qwen-VLA-Base.
Stage III: SFT (Supervised Fine-Tuning). Starting from the CPT checkpoint, we branch into two tracks: multi-task SFT jointly fine-tunes on manipulation, navigation, VQA, and spatial grounding; real-robot SFT fine-tunes on in-house teleoperation data for physical deployment.
Stage IV: RL (Reinforcement Learning). Starting from the SFT checkpoint, we use PPO to directly optimize closed-loop task success in simulation, producing the final model Qwen-VLA-Instruct. RL is conducted only in SimplerEnv, yet experiments show its gains transfer to unseen environments and robot embodiments.
A Single Generalist Model Can Match or Even Surpass Specialist Models#
The experimental results show the potential of Qwen-VLA as a generalist policy model. A single model can cover multiple manipulation benchmarks, including LIBERO, Simpler, RoboCasa, and RoboTwin, while approaching or surpassing specialized policy models on several tasks.
| Benchmark | Best Specialist Model | Qwen-VLA |
|---|---|---|
| LIBERO | ABot-M0 98.6% | 97.9% |
| RoboCasa-GR1 | ABot-M0 58.3% | 56.7% |
| Simpler-WidowX | StarVLA-OFT 64.6% | 73.7% |
| RoboTwin-Easy / Hard | ABot-M0 86.0% / 85.0% | 86.1% / 87.2% |
On robotic manipulation benchmarks, Qwen-VLA-Instruct achieves 97.9% on LIBERO, 73.7% on Simpler-WidowX, and 86.1% / 87.2% on RoboTwin-Easy / Hard. Many of the compared methods are specialist models fine-tuned for individual benchmarks, while Qwen-VLA is a unified generalist model trained under a single framework.
On vision-language navigation (VLN-CE), Qwen-VLA-Instruct achieves 69.0% Oracle Success Rate and 57.5% Success Rate on R2R Val-Unseen, and 59.6% SR and 47.8% SPL on the more challenging RxR Val-Unseen, surpassing all open-source baselines.
In real-world ALOHA dual-arm experiments, Qwen-VLA pretrained model achieves 83.6% average in-domain success and 76.9% average OOD success, substantially outperforming training from scratch (48.5% / 36.2%) and $\pi{0.5}$π 0.5 (71.6% / 41.5%).
Real-World Out-of-Distribution Generalization#
We also care about how Qwen-VLA generalizes on real robots.
In real-world ALOHA dual-arm robot experiments, Qwen-VLA demonstrates generalization to unseen colors, objects, backgrounds, positions, and language instructions. Compared with policies trained from scratch, models pretrained with Qwen-VLA show clear improvements under real-world out-of-distribution settings.
This part is best shown through videos. The following demonstrations are tested with the Qwen-VLA-Base model. When asked to “pick up the green ball” or “pick up the blue ball,” the model can correctly act based on color-specific instructions. When presented with unseen objects such as toys, vegetables, or sunglasses, it can still follow language commands to grasp or move them. When the background, lighting, and tabletop layout change, the model remains relatively stable. For compositional tasks such as “tidy up the table,” it can identify multiple targets and execute multi-step operations.
Compared with tables alone, these videos better illustrate the core value of Qwen-VLA:
The model is not merely memorizing action templates in a fixed environment. It is learning to understand goals and act under real-world variations.
Zero-Shot Generalization in Dynamic Scenes#
Beyond static tabletop manipulation, Qwen-VLA also shows zero-shot generalization in dynamic manipulation tasks.
On the DOMINO dynamic manipulation benchmark, Qwen-VLA-Instruct is not specifically fine-tuned for the benchmark, yet it still achieves a 26.6% success rate and a 39.5 manipulation score, outperforming a range of standard VLA baselines and even some specialist models for dynamic manipulation.
This suggests that the model is not only learning grasping templates in static scenes, but also acquiring a more transferable action prior from spatial understanding to motion control. Given visual observations, language goals, and its action generation capability, the model can directly produce coherent action sequences and complete tasks within dynamic interaction windows.
From Multimodal Understanding to Embodied Intelligence#
Qwen-VLA is a natural extension of Qwen’s multimodal capabilities toward embodied intelligence.
In the past, multimodal models mainly focused on understanding the world. With Qwen-VLA, we further explore how models can generate actions in the physical world based on vision and language.
Qwen-VLA unifies robotic manipulation, vision-language navigation, and cross-embodiment control. It connects Qwen’s visual understanding and spatial reasoning capabilities to continuous action generation. Through joint pretraining on real robot data, human egocentric data, synthetic simulation data, and general vision-language data, it learns more general embodied experience. It also demonstrates the potential of generalist policy models across manipulation benchmarks, real-world out-of-distribution generalization, and zero-shot dynamic manipulation.
Embodied intelligence is still at an early stage. Long-horizon real-world tasks, failure recovery, continual learning, and more complex human-robot-environment interactions remain challenging. But Qwen-VLA points to a clear next step:
Models should not only understand the world — they should also learn to act in it.
@article{qwenvla, title={Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments}, author={Qwen Team}, year={2026}, eprint={2605.30280}, archivePrefix={arXiv}, primaryClass={cs.RO}, url={https://arxiv.org/abs/2605.30280},}