Alibaba Tongyi Launches General Embodied Intelligence Unified Action Framework Qwen-VLA

Qwen-VLA: A Vision-Language-Action Base Unifying Manipulation, Navigation, and Trajectory Prediction in One Model

Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments Alibaba Qwen Team (authors include Qwen Lead Junyang Lin, as well as Dayiheng Liu, Shuai Bai, Jingren Zhou, etc.), 2026.05 · arXiv:2605.30280 Roadmap: Unified VLA Base · Qwen3.5-4B VL Backbone + 1.15B DiT Flow-Matching Action Decoder · Manipulation / Navigation / Trajectory Prediction integrated into a single "Action-Trajectory Prediction" Framework · Embodiment-aware prompt conditioning for cross-embodiment support

⚠️ High Credibility Note (Please Read First): arXiv ID 2605 corresponds to May 2026. This paper is extremely new (submitted 2026-05-28), not yet fully reviewed by the community, and has no independent third-party reproduction. All benchmark numbers and comparisons below are self-reported by the Qwen Team; details such as parameter split (Qwen3.5-4B backbone + 1.15B DiT decoder) come from the paper abstract, HuggingFace papers page, and official GitHub

QwenLM/Qwen-VLA README/Architecture Diagram. As of writing, the full technical report (arXiv HTML) has not been released. Some internal mechanisms can only be inferred from the abstract and README. Such instances are marked as "Details Not Public" in this article.

TL;DR

Qwen-VLA is a unified embodied base model launched by Alibaba's Qwen Team. Its core claim is simple: pack "manipulation, navigation, and trajectory prediction"—embodied decision problems previously handled separately—into a single VLA model, using one set of weights to handle all tasks—rather than training a specialized model for each task, environment, or robot embodiment.

Two pieces to achieve this:

- Architecture: On top of the Qwen3.5-4B Vision-Language backbone, it connects a 1.15B DiT (Diffusion Transformer) flow-matching action decoder, extending Qwen's original "perception—understanding—reasoning" capabilities to continuous action and trajectory generation. The VL backbone handles looking at images, reading instructions, and spatial reasoning; the DiT decoder takes noisy actions as input and iteratively denoises them into clean continuous action blocks/trajectories.

- Interface: Uses embodiment-aware prompt conditioning—writing "what kind of robot is current, what control convention is used" into the text prompt. Thus, one set of weights serves multiple platforms; switching embodiments only requires changing the prompt, without needing independent output heads for each robot.

Key Conclusions (all self-reported by authors): A unified generalist Qwen-VLA matches or exceeds "specialized models fine-tuned separately for each benchmark" on multiple simulation and real-world benchmarks. Authors use this to shift the embodied intelligence narrative from "skill specialists" to "generalist actors". Representative numbers: LIBERO 97.9%, Simpler-WidowX 73.7%, RoboTwin-Easy/Hard 86.1%/87.2%, R2R 69.0% OSR, RxR 59.6% SR, Real-world ALOHA OOD Average 76.9%, DOMINO Dynamic Manipulation Zero-shot 26.6%.

1. The Problem to Solve

Embodied intelligence has long been studied "sliced by task": one model for manipulation, one for navigation, another for trajectory prediction. The result is fragmented capabilities and poor generalization across tasks, environments, and robot embodiments—a manipulation policy trained well on one robotic arm fails when changing embodiment or scene.

Qwen-VLA aims to answer a direct question: Can these heterogeneous embodied decision problems be unified into a single Vision-Language-Action model? Specifically broken down into three layers:

- Cross-task unification: Manipulation (grasp/place, organize), Navigation (walk to target in room), Trajectory Prediction (predict future motion) are essentially different output forms (joint actions vs. paths vs. trajectories). Can a shared "action-trajectory prediction" framework express all of them so one model learns them simultaneously?

- Cross-embodiment unification: Different robots have different degrees of freedom, camera configurations, and control conventions. Can we avoid dedicated output heads for each platform and use a lightweight method (text prompt) to tell the model "who is current, how to control"?

- Cross-environment generalization (OOD): Real deployment encounters unseen scene layouts, backgrounds, lighting, object placements, and even robot embodiment changes. Can unified large-scale pre-training bring robust generalization to these **out-of-distribution** changes?

Qwen-VLA positions itself by using Qwen's VL modeling stack (perception, understanding, reasoning) as a base, extending downwards to continuous action and trajectory generation, thereby creating a unified embodied base.

2. Method and Architecture

Qwen-VLA structure summarized in one sentence: Qwen3.5-4B VL Backbone + DiT Flow Matching Action Decoder, plus embodiment-aware text prompts, feeding multi-task multi-embodiment data into the same set of weights. Below are three subsections: 2.1 Overall Dual-Stream Architecture, 2.2 Unified "Action-Trajectory Prediction" Framework and Embodiment-Aware Prompts, 2.3 Progressive Training Recipe and Data Mixing.

Figure Caption (Based on official GitHub QwenLM/Qwen-VLA architecture diagram qwenvla_overview.png): In the center is the Qwen3.5 Vision-Language Model backbone. Left input is Observed Images + Prompt (text prompt). Top right is the Diffusion Transformer (DiT) Action Decoder—internally stacked Self-Attention + AdaLN + Feed-Forward MLP modules, taking Noisy Action and Timestep (denoising timestep) as input, outputting Clean Action via an MLP. Top three robot cartoons labeled VLA / VLN / VL indicate the same model covers three capabilities. Bottom and right show Manipulation, Navigation, Vision-Language Understanding sample scenes, intuitively expressing "one model, multi-task, multi-environment, multi-embodiment".

2.1 Overall Architecture: VL Backbone + DiT Flow Matching Decoder (Dual-Stream)

Qwen-VLA follows the mainstream VLA dual-stream approach: using a strong VL backbone for semantic understanding and an independent continuous action expert for high-frequency control. Specific division:

- VL Backbone — Qwen3.5-4B: Consumes observed images (multi-view possible) + text prompt, outputs representations carrying scene semantics, spatial reasoning, and instruction understanding. This branch inherits from Qwen's vision-language modeling stack, responsible for "understanding the scene, reading instructions, figuring out which direction to move".

- Action Decoder — 1.15B DiT Flow Matching: This is the part that "translates" VL representations into continuous actions/trajectories. It is a Diffusion Transformer, core modules are Self-Attention + AdaLN (Adaptive Layer Normalization, conditioned on denoising timestep) + FFN, stacked layers. Training/inference uses flow matching: starting from Gaussian noise, conditioned on VL backbone representations, iteratively denoising to finally produce clean continuous action blocks or trajectories.

Relation to π0: This is a continuation of the "VL Backbone + Independent Flow Matching Action Expert Outputting Continuous Action Blocks" paradigm pioneered by the π0 series. The difference is Qwen-VLA's action expert is an explicit DiT (Diffusion Transformer) (same family as GR00T N1's System 1, DiT + Flow Matching), and extends output from "manipulation action blocks" to a unified space of "actions and trajectories" (see 2.2).

Note: Specific coupling between VL backbone and DiT decoder (how attention interacts, how representations are injected into DiT, whether backbone is shared/frozen, action block length and control frequency, etc.)—these engineering details are not public before the full technical report is released. This article does not speculate.

2.2 Unified "Action-Trajectory Prediction" Framework + Embodiment-aware Prompt

Qwen-VLA lands "unification" on two mechanisms:

(a) Unified action-and-trajectory prediction framework. The paper casts manipulation, navigation, first-person (egocentric) action modeling, and trajectory prediction all into a shared "action-trajectory prediction" space. Intuitively: whether it's robotic arm joint actions, mobile robot navigation paths, or prediction of a future trajectory, all can be expressed as "a target sequence in a certain continuous space, conditioned on observations and instructions, needing to be denoised and generated by the DiT decoder". Thus, heterogeneous tasks reuse the same decoder and same supervision, without needing a separate path planning head for navigation or action head for manipulation. The paper claims this brings shareable visual grounding, spatial reasoning, and continuous action generation capabilities across different task families.

(b) Embodiment-aware prompt conditioning. To support multiple robot platforms, Qwen-VLA does not equip independent output heads for each platform, but writes "what kind of embodiment is current, what control convention is adopted" as a robot-specific text description, fed as part of the prompt. Thus: one set of weights serves all platforms, switching embodiments only requires changing a text prompt. This is the key move for Qwen-VLA to achieve "cross-embodiment unification", and the core difference from the "one fine-tuned model per robot" route.

2.3 Progressive Training Recipe and Heterogeneous Data Mixing

Qwen-VLA uses a progressive training recipe, jointly pre-training on large-scale heterogeneous embodied data, data.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.