{ "@context": "https://schema.org", "@id": "https://aggr.aymericbeaumet.com/items/qwen-ai/2026-06-16-qwen-robotworld-boundless-worlds-for-embodied-agents/#webpage", "@type": [ "WebPage", "ArchiveComponent" ], "articleBody": "![Qwen-RobotNav banner](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/suite/singleworld.png)\n\n![Qwen-RobotWorld](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotworld/hero_figure.png)\n\nEmbodied intelligence requires agents to perceive, reason about, and act within physical environments. World models offer a scalable path forward — but current approaches face a fundamental tension. *General video generation models* learn rich visual priors but **lack the ability to model embodied physics**. *Domain-specific embodied models* are tailored to individual scenarios and **cannot generalize across embodiments**.\n\n**Qwen-RobotWorld bridges this gap by treating natural language as a universal action interface.** A single instruction like *\"pick up the red cup and place it on the shelf\"* implicitly encodes the complete action sequence, goal state, and physical constraints — **no robot-specific control interface needed**. This allows manipulation, autonomous driving, and indoor navigation to be trained jointly, with each domain's physical knowledge reinforcing the others.\n\n> **Language unifies the action space**: world knowledge and embodied knowledge reinforce each other within a single model, enabling **cross-scenario, cross-task physical generalization**.\n\n* * *\n\n## Key Highlights\n\nTop-Tier Across 4 Benchmarks\n\n20+ Robot Embodiments Unified\n\n8.6M Cross-Scenario Training Pairs\n\n1300+ Manipulation Skills\n\n- **Language-Driven Unified Action Interface** — natural language standardizes **20+ robot embodiments** and **500+ action categories** into one interface, enabling joint cross-scenario training\n- **Dual-Stream Diffusion World Model** — MMDiT with **Qwen2.5-VL** as action encoder, combining deep language understanding with internalized physical world knowledge\n- **Cross-Scenario Physical Generalization** — manipulation, driving, navigation, and human-to-robot transfer jointly trained under **8.6M video-text pairs**\n- **Multi-View Geometrically Consistent Generation** — synchronized **2–4 camera streams** with 3D-consistent object identity and motion trajectories\n\n* * *\n\n## Model Architecture\n\n### Dual-Stream Diffusion World Model\n\n![Qwen-RobotWorld Architecture](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotworld/figure_model_structure.png)\n\nQwen-RobotWorld adopts a dual-stream Multimodal Diffusion Transformer (MMDiT):\n\n- **Understanding stream** processes semantic features from a frozen Qwen2.5-VL encoder, representing the language action $a\\_t$.\n- **Generation stream** processes visual latents from a video-compatible VAE, representing the visual state $s\\_t$.\n\nThe two streams interact via joint attention at every layer, enabling bidirectional cross-modal fusion throughout the denoising process.\n\nUsing an MLLM as the action encoder — rather than lightweight encoders like T5 or CLIP — provides two key advantages: (1) deep language understanding accurately parses complex, compositional instructions into precise condition signals; (2) internalized world knowledge (e.g., that robot arms are rigid bodies with fixed joint constraints) implicitly constrains physically plausible transitions, preventing common failure modes like object deformation across frames.\n\n### Scene2Robot: Human-to-Robot Transfer\n\n**Scene2Robot** enables cross-embodiment video editing: human demonstrations are retargeted to **14 robot morphologies** via a multi-segment conditioning mechanism, where joint attention allows the generation to simultaneously attend to scene appearance and robot motion trajectory. This capability both serves as a data scaling engine during training and enables human-to-robot transfer at inference time.\n\n![Scene2Robot Architecture](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotworld/figure_scene2robot_architecture.png)\n\n### Multi-View Geometrically Consistent Generation\n\nSingle-camera observation inevitably occludes critical contact and spatial details. Qwen-RobotWorld generates 2–4 synchronized camera streams — main view, wrist-mounted views, and third-person views — with geometrically consistent object identity and motion across all viewpoints. During training, synchronized frames from multiple cameras are spatially concatenated into a single input; the model generates all views simultaneously, with **asymmetric 3D RoPE** providing spatial encoding and attention layers naturally establishing cross-view correspondence — without any architectural modification. This cross-view consistency further acts as a geometric regularizer, teaching the model object shape, depth, and spatial layout.\n\n* * *\n\n## Data: Embodied World Knowledge\n\n### EWK Dataset\n\n![Embodied World Knowledge Dataset](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotworld/figure_dataset_visual.png)\n\nThe Embodied World Knowledge (EWK) dataset is organized along four complementary axes, each targeting a distinct source of physical variation:\n\n- **Multi-Embodiment** — human hands, 7 robot arm configurations, ego vehicles, mobile agents, spanning 20+ distinct robot models\n- **Multi-Task** — atomic manipulation skills, long-horizon compositions, locomotion, dynamic/deformable interactions across 500+ action categories\n- **Multi-Scenario** — real-world first, sim-augmented: kitchens, workshops, outdoor settings, plus photorealistic simulation for downstream VLA evaluation\n- **Multi-View** — main, wrist, and synchronized multi-view streams (~1.6M of 6M embodied samples include 2–4 view concatenations)\n\nView Detailed Dataset Inventory\n\n| Dataset | Embodiment | Views | Contribution |\n| ----------------------------------------------- | -------------------------------- | ------------------------ | --------------------------------------------- |\n| Manipulation (~5.9M samples) | | | |\n| EgoHOD, EPIC-Kitchens, Egocentric-10k | Human hands | Egocentric | Dexterity & coordination prior |\n| Bridge V2, RH20T, Droid | Single-arm grippers | 3rd-person + wrist | Interaction primitives |\n| Robomind, RoboCoin | Single/dual-arm, humanoids | Ego + 3rd-person | Cross-embodiment generalization |\n| Agibot-World, Galaxea | Single-arm (gripper + dexterous) | Synced ego + wrist + 3rd | Temporal & multi-view consistency |\n| Qwen-Aloha (internal) | Dual-arm grippers | Head + dual wrist | Multi-view grasping prior |\n| ActionNet, OpenLoong | Dexterous hands | Wrist + 3rd-person | Fine-grained dexterity |\n| Autonomous Driving (~200K samples) | | | |\n| Waymo, NVIDIA PhysicalAI-AD, Bench2Drive, Sekai | Ego vehicle | Surround-view | Large-scale ego-motion & multi-agent dynamics |\n| Indoor Navigation (6K+ episodes) | | | |\n| VLNVerse | Mobile agent | Egocentric | Room-scale spatial reasoning |\n| Human-to-Robot Transfer | | | |\n| Scene2Robot (synthesized) | 14 robot morphologies | Multi-view | Cross-embodiment video editing |\n\n### Action-Language Mapping\n\nThe central challenge in building a universal world model is representational heterogeneity: manipulation uses joint angles, driving uses steering commands, navigation uses heading vectors — each requiring a separate model. Our **action-language mapping framework** resolves this by projecting all action signals onto a shared natural language space, so that videos from a Franka gripper, an autonomous vehicle, and a navigation agent all become instances of the same language-conditioned video generation task.\n\nA **hierarchical five-layer annotation pipeline** ensures caption quality and precision:\n\n1\n\n**Task Goal** High-level intent — what should change between states\n\n2\n\n**Action Detail** Spatio-temporal trajectories with explicit viewpoint declaration\n\n3\n\n**Physical Feedback** Observable consequences on the environment\n\n4\n\n**Comprehensive Caption** Full description for precise prediction\n\n5\n\n**Concise Caption** Essential elements for brief task-level commands\n\nDuring training, comprehensive and concise descriptions are sampled with equal probability, so the model handles both detailed trajectory specifications and brief task-level commands.\n\n* * *\n\n## Training\n\nTraining follows a **general-to-expert progressive curriculum**:\n\n| Stage | Phase | Data Mix | Objective |\n| --------------------------------- | --------------------------------- | --------------------------------- | -------------------------------- |\n| **Pretraining** | T2I / T2V / TI2V joint | General data | Build foundational visual priors |\n| Human interaction | Ego4D, EPIC-Kitchen, etc. | Grasping & tool-use priors | |\n| **SFT** | Phase 1: Single-view manipulation | Embodied + general joint training | Core manipulation physics |\n| Phase 2: Multi-view expansion | Broaden viewpoint coverage | | |\n| Phase 3: Multi-view concatenation | Cross-view geometric consistency | | |\n| Phase 4: Complex cross-domain | Long-horizon & cross-scenario | | |\n\nPretraining on general data and human interaction videos (Ego4D, EPIC-Kitchen) builds broad visual priors — the T2I task specifically anchors object geometry that transfers to video generation through the shared backbone. SFT then progressively deepens embodied expertise across four phases while keeping general data in every batch, ensuring both capabilities advance together rather than trade off.\n\n* * *\n\n## Demos\n\n### Fine-Grained Language Grounding\n\nGiven identical initial frames, the model produces qualitatively distinct videos when a single keyword differs. It also handles complex, multi-step instructions requiring long-horizon reasoning.\n\nContrastive Instruction Following:\n\nL: Pick up the red strawberry\\\nR: Pick up the yellow potato\n\nL: Place pen on the wooden tray\\\nR: Place pen on the white paper\n\nL: Hand glue to the person\\\nR: Place glue into the penholder\n\nComplex Multi-Step Instructions:\n\nSequentially pick up the red and yellow bell peppers, place them on the table from left to right\n\nGrab the yellow-and-blue stacked block, position it above the green-and-blue stacked block\n\n### Cross-Domain Generalization\n\n(A) Cross-Embodiment:\n\n(B) Cross-Task × Cross-Environment:\n\n(C) Multi-View Consistency:\n\n(D) Zero-Shot Robustness:\n\nOurs\n\nLVP\n\nCosmos2.5-14B\n\n### Human-to-Robot Transfer\n\nThe Scene2Robot mechanism preserves task intent from a human demonstration (left) while adapting motion to embodiment-specific kinematic constraints (right).\n\n### Beyond Manipulation: Driving and Navigation\n\nThe learned world model generalizes beyond robot manipulation to broader mobility scenarios.\n\n**Autonomous Driving.** Generated driving episodes from Bench2Drive, NVIDIA PhysicalAI-AD, Sekai, and Waymo demonstrate coherent scene dynamics and vehicle behaviors.\n\n**Indoor Navigation.** Egocentric navigation episodes from VLNVerse show the model’s ability to simulate first-person movement through complex indoor environments.\n\n> From tabletop manipulation to autonomous driving and indoor navigation — Qwen-RobotWorld demonstrates that a unified world model can generalize beyond a single morphology or scenario family.\n\n* * *\n\n## Performance\n\nWe evaluate against general video generation models (Sora2, Veo3, Wan2.6, Kling, LTX-2) and embodied world models (Cosmos, LVP, GigaWorld, Vidar, Wow) across four benchmarks.\n\n![Benchmark Radar Overview](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotworld/radar_figure.png)\n\nEWMBench\n\n4.60\n\nEmbodied motion fidelity\n\nDreamGen\n\n4.952\n\nInstruction following & physics alignment\n\nWorldModelBench\n\n8.99\n\nPhysical reasoning & instruction following\n\nPBench\n\n0.804\n\nPhysical behavior evaluation\n\nEWMBench\n\n| Type | Model | SceneC | HSD | Dyn | nDTW | Diversity | BLEU | CLIP | Logics | Overall |\n| --------- | -------- | --------- | --------- | --------- | --------- | --------- | ----- | ----- | --------- | -------- |\n| General | Veo3 | 0.842 | 0.213 | 0.193 | 0.161 | 0.022 | 0.214 | 0.897 | 0.947 | 3.49 |\n| Wan2.6 | 0.671 | 0.203 | 0.090 | 0.172 | 0.050 | 0.162 | 0.874 | 1.000 | 3.22 | |\n| Kling | 0.821 | 0.327 | 0.182 | 0.342 | 0.017 | 0.259 | 0.901 | 1.000 | 3.85 | |\n| LTX-2 | 0.785 | 0.208 | 0.128 | 0.244 | 0.012 | 0.143 | 0.887 | 0.500 | 2.91 | |\n| Sora2 | 0.853 | 0.281 | 0.349 | 0.275 | 0.031 | 0.247 | 0.910 | 0.947 | 3.89 | |\n| Embodied | Cosmos | 0.796 | 0.250 | 0.205 | 0.253 | 0.080 | 0.123 | 0.846 | 0.733 | 3.29 |\n| GigaWorld | 0.871 | 0.305 | 0.085 | 0.278 | 0.028 | 0.205 | 0.887 | 0.900 | 3.56 | |\n| LVP | 0.880 | 0.425 | 0.043 | 0.623 | 0.009 | 0.218 | 0.900 | 0.952 | 4.05 | |\n| Vidar | 0.734 | 0.188 | 0.152 | 0.177 | 0.065 | 0.161 | 0.882 | 0.941 | 3.30 | |\n| Wow | 0.887 | 0.249 | 0.053 | 0.257 | 0.027 | 0.193 | 0.900 | 0.952 | 3.52 | |\n| | **Ours** | **0.914** | **0.566** | **0.343** | **0.671** | 0.011 | 0.208 | 0.883 | **1.000** | **4.60** |\n\nStrong motion fidelity (HSD **0.566**), high scene consistency (**0.914**), and perfect logic constraint satisfaction.\n\nDreamGen Bench\n\n| Model | GR1-Env PA | GR1-Env IF | GR1-Object PA | GR1-Object IF | GR1-Behavior PA | GR1-Behavior IF | Total |\n| ---------- | ---------- | ---------- | ------------- | ------------- | --------------- | --------------- | --------- |\n| Cosmos-sft | 0.709 | 0.655 | 0.775 | 0.720 | 0.649 | 0.621 | 4.129 |\n| LVP | 0.810 | 0.772 | 0.745 | 0.829 | 0.713 | 0.889 | 4.758 |\n| Vidar | 0.445 | 0.647 | 0.478 | 0.726 | 0.394 | 0.651 | 3.341 |\n| GigaWorld | 0.621 | 0.933 | 0.500 | 0.852 | 0.426 | 0.884 | 4.216 |\n| Wow | 0.793 | 0.826 | 0.755 | 0.849 | 0.809 | 0.696 | 4.728 |\n| **Ours** | **0.828** | 0.793 | **0.840** | **0.878** | 0.781 | 0.832 | **4.952** |\n\nStrong object-level compositional generalization (GR1-Object IF: **0.878**), with consistent physics alignment across all subsets.\n\nWorldModelBench\n\n| Type | Model | Instr. (0-3) | Frame | Temp | Newton | Mass | Fluid | Penetr. | Grav. | Phys. | Total |\n| --------- | -------- | ------------ | ----- | ---- | -------- | -------- | -------- | ------- | -------- | -------- | -------- |\n| General | Veo3 | 2.52 | 0.98 | 0.95 | 1.00 | 0.89 | 0.99 | 0.91 | 1.00 | 4.80 | 9.25 |\n| Wan2.6 | 2.50 | 0.99 | 0.95 | 1.00 | 0.89 | 0.99 | 0.94 | 1.00 | 4.83 | 9.27 | |\n| Sora2 | 2.21 | 0.96 | 0.93 | 1.00 | 0.91 | 0.99 | 0.95 | 1.00 | 4.84 | 8.93 | |\n| Kling | 1.59 | 0.97 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 5.00 | 8.55 | |\n| LTX-2 | 1.97 | 0.69 | 0.62 | 0.99 | 0.60 | 1.00 | 0.73 | 1.00 | 4.32 | 7.61 | |\n| Embodied | Cosmos | 2.14 | 1.00 | 0.94 | 1.00 | 0.92 | 1.00 | 0.94 | 1.00 | 4.86 | 8.94 |\n| LVP | 2.01 | 0.89 | 0.91 | 1.00 | 0.93 | 0.99 | 0.95 | 1.00 | 4.87 | 8.67 | |\n| GigaWorld | 2.13 | 0.59 | 0.46 | 1.00 | 0.48 | 0.99 | 0.69 | 0.98 | 4.13 | 7.31 | |\n| Vidar | 1.62 | 0.54 | 0.45 | 1.00 | 0.56 | 1.00 | 0.85 | 1.00 | 4.40 | 7.01 | |\n| Wow | 2.05 | 0.76 | 0.65 | 1.00 | 0.65 | 0.99 | 0.81 | 1.00 | 4.45 | 7.91 | |\n| | **Ours** | **2.33** | 0.87 | 0.85 | **1.00** | **1.00** | **1.00** | 0.94 | **1.00** | **4.94** | **8.99** |\n\n**Perfect physics adherence (1.00)** across Newton's laws, mass conservation, fluid dynamics, and gravity, with strong instruction following (2.33/3.0).\n\nPBench\n\n| Type | Model | I2V-Bg | I2V-S | Aes | Img | Bg-Con | Mot | Sub-Con | O-Con | Quality | Domain | Overall |\n| --------- | -------- | ------ | ----- | ----- | ----- | --------- | --------- | ------- | ----- | ------- | --------- | --------- |\n| General | Veo3 | 0.975 | 0.980 | 0.526 | 0.698 | 0.938 | 0.994 | 0.927 | 0.128 | 0.771 | 0.882 | **0.827** |\n| Wan2.6 | 0.856 | 0.843 | 0.514 | 0.719 | 0.906 | 0.978 | 0.843 | 0.136 | 0.724 | 0.832 | 0.778 | |\n| Sora2 | 0.981 | 0.973 | 0.487 | 0.672 | 0.961 | 0.994 | 0.954 | 0.129 | 0.769 | 0.841 | 0.805 | |\n| Kling | 0.982 | 0.979 | 0.521 | 0.699 | 0.920 | 0.990 | 0.927 | 0.124 | 0.768 | 0.874 | 0.821 | |\n| LTX-2 | 0.948 | 0.955 | 0.506 | 0.622 | 0.932 | 0.986 | 0.904 | 0.118 | 0.746 | 0.845 | 0.796 | |\n| Embodied | LVP | 0.979 | 0.981 | 0.515 | 0.679 | 0.954 | 0.991 | 0.962 | 0.116 | 0.772 | 0.812 | 0.792 |\n| GigaWorld | 0.957 | 0.944 | 0.495 | 0.641 | 0.925 | 0.984 | 0.892 | 0.128 | 0.746 | 0.841 | 0.794 | |\n| Wow | 0.967 | 0.957 | 0.517 | 0.689 | 0.941 | 0.980 | 0.929 | 0.111 | 0.761 | 0.786 | 0.774 | |\n| Vidar | 0.935 | 0.922 | 0.501 | 0.573 | 0.912 | 0.982 | 0.863 | 0.120 | 0.726 | 0.810 | 0.768 | |\n| Cosmos | 0.974 | 0.973 | 0.470 | 0.663 | 0.940 | 0.989 | 0.931 | 0.160 | 0.763 | 0.840 | 0.802 | |\n| | **Ours** | 0.956 | 0.943 | 0.455 | 0.649 | **0.956** | **0.990** | 0.933 | 0.124 | 0.751 | **0.857** | **0.804** |\n\nStrong domain understanding (**0.857**) and motion smoothness (**0.990**), reflecting consistent temporal coherence across physical scenarios.\n\n[← Back to Qwen-Robot Suite](https://qwen.ai/blog?id=qwen-robotsuite)\n\n* * *\n\n## Citation\n\nIf you find our work helpful, feel free to give us a cite.\n\n```bibtex\n@article{qwenrobot-world,\n title={Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation},\n author={Qwen Team},\n year={2026}\n}\n```\n", "authors": [ "QwenTeam" ], "captured_at": "2026-10-19T01:26:53.998049939+00:00", "category": "ai", "content_html": "

\"Qwen-RobotNav

\n

\"Qwen-RobotWorld\"

\n

Embodied intelligence requires agents to perceive, reason about, and act within physical environments. World models offer a scalable path forward — but current approaches face a fundamental tension. General video generation models learn rich visual priors but lack the ability to model embodied physics. Domain-specific embodied models are tailored to individual scenarios and cannot generalize across embodiments.

\n

Qwen-RobotWorld bridges this gap by treating natural language as a universal action interface. A single instruction like "pick up the red cup and place it on the shelf" implicitly encodes the complete action sequence, goal state, and physical constraints — no robot-specific control interface needed. This allows manipulation, autonomous driving, and indoor navigation to be trained jointly, with each domain's physical knowledge reinforcing the others.

\n
\n

Language unifies the action space: world knowledge and embodied knowledge reinforce each other within a single model, enabling cross-scenario, cross-task physical generalization.

\n
\n
\n

Key Highlights

\n

Top-Tier Across 4 Benchmarks

\n

20+ Robot Embodiments Unified

\n

8.6M Cross-Scenario Training Pairs

\n

1300+ Manipulation Skills

\n\n
\n

Model Architecture

\n

Dual-Stream Diffusion World Model

\n

\"Qwen-RobotWorld

\n

Qwen-RobotWorld adopts a dual-stream Multimodal Diffusion Transformer (MMDiT):

\n\n

The two streams interact via joint attention at every layer, enabling bidirectional cross-modal fusion throughout the denoising process.

\n

Using an MLLM as the action encoder — rather than lightweight encoders like T5 or CLIP — provides two key advantages: (1) deep language understanding accurately parses complex, compositional instructions into precise condition signals; (2) internalized world knowledge (e.g., that robot arms are rigid bodies with fixed joint constraints) implicitly constrains physically plausible transitions, preventing common failure modes like object deformation across frames.

\n

Scene2Robot: Human-to-Robot Transfer

\n

Scene2Robot enables cross-embodiment video editing: human demonstrations are retargeted to 14 robot morphologies via a multi-segment conditioning mechanism, where joint attention allows the generation to simultaneously attend to scene appearance and robot motion trajectory. This capability both serves as a data scaling engine during training and enables human-to-robot transfer at inference time.

\n

\"Scene2Robot

\n

Multi-View Geometrically Consistent Generation

\n

Single-camera observation inevitably occludes critical contact and spatial details. Qwen-RobotWorld generates 2–4 synchronized camera streams — main view, wrist-mounted views, and third-person views — with geometrically consistent object identity and motion across all viewpoints. During training, synchronized frames from multiple cameras are spatially concatenated into a single input; the model generates all views simultaneously, with asymmetric 3D RoPE providing spatial encoding and attention layers naturally establishing cross-view correspondence — without any architectural modification. This cross-view consistency further acts as a geometric regularizer, teaching the model object shape, depth, and spatial layout.

\n
\n

Data: Embodied World Knowledge

\n

EWK Dataset

\n

\"Embodied

\n

The Embodied World Knowledge (EWK) dataset is organized along four complementary axes, each targeting a distinct source of physical variation:

\n\n

View Detailed Dataset Inventory

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
DatasetEmbodimentViewsContribution
Manipulation (~5.9M samples)
EgoHOD, EPIC-Kitchens, Egocentric-10kHuman handsEgocentricDexterity & coordination prior
Bridge V2, RH20T, DroidSingle-arm grippers3rd-person + wristInteraction primitives
Robomind, RoboCoinSingle/dual-arm, humanoidsEgo + 3rd-personCross-embodiment generalization
Agibot-World, GalaxeaSingle-arm (gripper + dexterous)Synced ego + wrist + 3rdTemporal & multi-view consistency
Qwen-Aloha (internal)Dual-arm grippersHead + dual wristMulti-view grasping prior
ActionNet, OpenLoongDexterous handsWrist + 3rd-personFine-grained dexterity
Autonomous Driving (~200K samples)
Waymo, NVIDIA PhysicalAI-AD, Bench2Drive, SekaiEgo vehicleSurround-viewLarge-scale ego-motion & multi-agent dynamics
Indoor Navigation (6K+ episodes)
VLNVerseMobile agentEgocentricRoom-scale spatial reasoning
Human-to-Robot Transfer
Scene2Robot (synthesized)14 robot morphologiesMulti-viewCross-embodiment video editing
\n

Action-Language Mapping

\n

The central challenge in building a universal world model is representational heterogeneity: manipulation uses joint angles, driving uses steering commands, navigation uses heading vectors — each requiring a separate model. Our action-language mapping framework resolves this by projecting all action signals onto a shared natural language space, so that videos from a Franka gripper, an autonomous vehicle, and a navigation agent all become instances of the same language-conditioned video generation task.

\n

A hierarchical five-layer annotation pipeline ensures caption quality and precision:

\n

1

\n

Task Goal High-level intent — what should change between states

\n

2

\n

Action Detail Spatio-temporal trajectories with explicit viewpoint declaration

\n

3

\n

Physical Feedback Observable consequences on the environment

\n

4

\n

Comprehensive Caption Full description for precise prediction

\n

5

\n

Concise Caption Essential elements for brief task-level commands

\n

During training, comprehensive and concise descriptions are sampled with equal probability, so the model handles both detailed trajectory specifications and brief task-level commands.

\n
\n

Training

\n

Training follows a general-to-expert progressive curriculum:

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
StagePhaseData MixObjective
PretrainingT2I / T2V / TI2V jointGeneral dataBuild foundational visual priors
Human interactionEgo4D, EPIC-Kitchen, etc.Grasping & tool-use priors
SFTPhase 1: Single-view manipulationEmbodied + general joint trainingCore manipulation physics
Phase 2: Multi-view expansionBroaden viewpoint coverage
Phase 3: Multi-view concatenationCross-view geometric consistency
Phase 4: Complex cross-domainLong-horizon & cross-scenario
\n

Pretraining on general data and human interaction videos (Ego4D, EPIC-Kitchen) builds broad visual priors — the T2I task specifically anchors object geometry that transfers to video generation through the shared backbone. SFT then progressively deepens embodied expertise across four phases while keeping general data in every batch, ensuring both capabilities advance together rather than trade off.

\n
\n

Demos

\n

Fine-Grained Language Grounding

\n

Given identical initial frames, the model produces qualitatively distinct videos when a single keyword differs. It also handles complex, multi-step instructions requiring long-horizon reasoning.

\n

Contrastive Instruction Following:

\n

L: Pick up the red strawberry
\nR: Pick up the yellow potato

\n

L: Place pen on the wooden tray
\nR: Place pen on the white paper

\n

L: Hand glue to the person
\nR: Place glue into the penholder

\n

Complex Multi-Step Instructions:

\n

Sequentially pick up the red and yellow bell peppers, place them on the table from left to right

\n

Grab the yellow-and-blue stacked block, position it above the green-and-blue stacked block

\n

Cross-Domain Generalization

\n

(A) Cross-Embodiment:

\n

(B) Cross-Task × Cross-Environment:

\n

(C) Multi-View Consistency:

\n

(D) Zero-Shot Robustness:

\n

Ours

\n

LVP

\n

Cosmos2.5-14B

\n

Human-to-Robot Transfer

\n

The Scene2Robot mechanism preserves task intent from a human demonstration (left) while adapting motion to embodiment-specific kinematic constraints (right).

\n

Beyond Manipulation: Driving and Navigation

\n

The learned world model generalizes beyond robot manipulation to broader mobility scenarios.

\n

Autonomous Driving. Generated driving episodes from Bench2Drive, NVIDIA PhysicalAI-AD, Sekai, and Waymo demonstrate coherent scene dynamics and vehicle behaviors.

\n

Indoor Navigation. Egocentric navigation episodes from VLNVerse show the model’s ability to simulate first-person movement through complex indoor environments.

\n
\n

From tabletop manipulation to autonomous driving and indoor navigation — Qwen-RobotWorld demonstrates that a unified world model can generalize beyond a single morphology or scenario family.

\n
\n
\n

Performance

\n

We evaluate against general video generation models (Sora2, Veo3, Wan2.6, Kling, LTX-2) and embodied world models (Cosmos, LVP, GigaWorld, Vidar, Wow) across four benchmarks.

\n

\"Benchmark

\n

EWMBench

\n

4.60

\n

Embodied motion fidelity

\n

DreamGen

\n

4.952

\n

Instruction following & physics alignment

\n

WorldModelBench

\n

8.99

\n

Physical reasoning & instruction following

\n

PBench

\n

0.804

\n

Physical behavior evaluation

\n

EWMBench

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
TypeModelSceneCHSDDynnDTWDiversityBLEUCLIPLogicsOverall
GeneralVeo30.8420.2130.1930.1610.0220.2140.8970.9473.49
Wan2.60.6710.2030.0900.1720.0500.1620.8741.0003.22
Kling0.8210.3270.1820.3420.0170.2590.9011.0003.85
LTX-20.7850.2080.1280.2440.0120.1430.8870.5002.91
Sora20.8530.2810.3490.2750.0310.2470.9100.9473.89
EmbodiedCosmos0.7960.2500.2050.2530.0800.1230.8460.7333.29
GigaWorld0.8710.3050.0850.2780.0280.2050.8870.9003.56
LVP0.8800.4250.0430.6230.0090.2180.9000.9524.05
Vidar0.7340.1880.1520.1770.0650.1610.8820.9413.30
Wow0.8870.2490.0530.2570.0270.1930.9000.9523.52
Ours0.9140.5660.3430.6710.0110.2080.8831.0004.60
\n

Strong motion fidelity (HSD 0.566), high scene consistency (0.914), and perfect logic constraint satisfaction.

\n

DreamGen Bench

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
ModelGR1-Env PAGR1-Env IFGR1-Object PAGR1-Object IFGR1-Behavior PAGR1-Behavior IFTotal
Cosmos-sft0.7090.6550.7750.7200.6490.6214.129
LVP0.8100.7720.7450.8290.7130.8894.758
Vidar0.4450.6470.4780.7260.3940.6513.341
GigaWorld0.6210.9330.5000.8520.4260.8844.216
Wow0.7930.8260.7550.8490.8090.6964.728
Ours0.8280.7930.8400.8780.7810.8324.952
\n

Strong object-level compositional generalization (GR1-Object IF: 0.878), with consistent physics alignment across all subsets.

\n

WorldModelBench

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
TypeModelInstr. (0-3)FrameTempNewtonMassFluidPenetr.Grav.Phys.Total
GeneralVeo32.520.980.951.000.890.990.911.004.809.25
Wan2.62.500.990.951.000.890.990.941.004.839.27
Sora22.210.960.931.000.910.990.951.004.848.93
Kling1.590.971.001.001.001.001.001.005.008.55
LTX-21.970.690.620.990.601.000.731.004.327.61
EmbodiedCosmos2.141.000.941.000.921.000.941.004.868.94
LVP2.010.890.911.000.930.990.951.004.878.67
GigaWorld2.130.590.461.000.480.990.690.984.137.31
Vidar1.620.540.451.000.561.000.851.004.407.01
Wow2.050.760.651.000.650.990.811.004.457.91
Ours2.330.870.851.001.001.000.941.004.948.99
\n

Perfect physics adherence (1.00) across Newton's laws, mass conservation, fluid dynamics, and gravity, with strong instruction following (2.33/3.0).

\n

PBench

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
TypeModelI2V-BgI2V-SAesImgBg-ConMotSub-ConO-ConQualityDomainOverall
GeneralVeo30.9750.9800.5260.6980.9380.9940.9270.1280.7710.8820.827
Wan2.60.8560.8430.5140.7190.9060.9780.8430.1360.7240.8320.778
Sora20.9810.9730.4870.6720.9610.9940.9540.1290.7690.8410.805
Kling0.9820.9790.5210.6990.9200.9900.9270.1240.7680.8740.821
LTX-20.9480.9550.5060.6220.9320.9860.9040.1180.7460.8450.796
EmbodiedLVP0.9790.9810.5150.6790.9540.9910.9620.1160.7720.8120.792
GigaWorld0.9570.9440.4950.6410.9250.9840.8920.1280.7460.8410.794
Wow0.9670.9570.5170.6890.9410.9800.9290.1110.7610.7860.774
Vidar0.9350.9220.5010.5730.9120.9820.8630.1200.7260.8100.768
Cosmos0.9740.9730.4700.6630.9400.9890.9310.1600.7630.8400.802
Ours0.9560.9430.4550.6490.9560.9900.9330.1240.7510.8570.804
\n

Strong domain understanding (0.857) and motion smoothness (0.990), reflecting consistent temporal coherence across physical scenarios.

\n

← Back to Qwen-Robot Suite

\n
\n

Citation

\n

If you find our work helpful, feel free to give us a cite.

\n
@article{qwenrobot-world,\n  title={Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation},\n  author={Qwen Team},\n  year={2026}\n}\n
\n", "content_markdown": "![Qwen-RobotNav banner](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/suite/singleworld.png)\n\n![Qwen-RobotWorld](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotworld/hero_figure.png)\n\nEmbodied intelligence requires agents to perceive, reason about, and act within physical environments. World models offer a scalable path forward — but current approaches face a fundamental tension. *General video generation models* learn rich visual priors but **lack the ability to model embodied physics**. *Domain-specific embodied models* are tailored to individual scenarios and **cannot generalize across embodiments**.\n\n**Qwen-RobotWorld bridges this gap by treating natural language as a universal action interface.** A single instruction like *\"pick up the red cup and place it on the shelf\"* implicitly encodes the complete action sequence, goal state, and physical constraints — **no robot-specific control interface needed**. This allows manipulation, autonomous driving, and indoor navigation to be trained jointly, with each domain's physical knowledge reinforcing the others.\n\n> **Language unifies the action space**: world knowledge and embodied knowledge reinforce each other within a single model, enabling **cross-scenario, cross-task physical generalization**.\n\n* * *\n\n## Key Highlights\n\nTop-Tier Across 4 Benchmarks\n\n20+ Robot Embodiments Unified\n\n8.6M Cross-Scenario Training Pairs\n\n1300+ Manipulation Skills\n\n- **Language-Driven Unified Action Interface** — natural language standardizes **20+ robot embodiments** and **500+ action categories** into one interface, enabling joint cross-scenario training\n- **Dual-Stream Diffusion World Model** — MMDiT with **Qwen2.5-VL** as action encoder, combining deep language understanding with internalized physical world knowledge\n- **Cross-Scenario Physical Generalization** — manipulation, driving, navigation, and human-to-robot transfer jointly trained under **8.6M video-text pairs**\n- **Multi-View Geometrically Consistent Generation** — synchronized **2–4 camera streams** with 3D-consistent object identity and motion trajectories\n\n* * *\n\n## Model Architecture\n\n### Dual-Stream Diffusion World Model\n\n![Qwen-RobotWorld Architecture](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotworld/figure_model_structure.png)\n\nQwen-RobotWorld adopts a dual-stream Multimodal Diffusion Transformer (MMDiT):\n\n- **Understanding stream** processes semantic features from a frozen Qwen2.5-VL encoder, representing the language action $a\\_t$.\n- **Generation stream** processes visual latents from a video-compatible VAE, representing the visual state $s\\_t$.\n\nThe two streams interact via joint attention at every layer, enabling bidirectional cross-modal fusion throughout the denoising process.\n\nUsing an MLLM as the action encoder — rather than lightweight encoders like T5 or CLIP — provides two key advantages: (1) deep language understanding accurately parses complex, compositional instructions into precise condition signals; (2) internalized world knowledge (e.g., that robot arms are rigid bodies with fixed joint constraints) implicitly constrains physically plausible transitions, preventing common failure modes like object deformation across frames.\n\n### Scene2Robot: Human-to-Robot Transfer\n\n**Scene2Robot** enables cross-embodiment video editing: human demonstrations are retargeted to **14 robot morphologies** via a multi-segment conditioning mechanism, where joint attention allows the generation to simultaneously attend to scene appearance and robot motion trajectory. This capability both serves as a data scaling engine during training and enables human-to-robot transfer at inference time.\n\n![Scene2Robot Architecture](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotworld/figure_scene2robot_architecture.png)\n\n### Multi-View Geometrically Consistent Generation\n\nSingle-camera observation inevitably occludes critical contact and spatial details. Qwen-RobotWorld generates 2–4 synchronized camera streams — main view, wrist-mounted views, and third-person views — with geometrically consistent object identity and motion across all viewpoints. During training, synchronized frames from multiple cameras are spatially concatenated into a single input; the model generates all views simultaneously, with **asymmetric 3D RoPE** providing spatial encoding and attention layers naturally establishing cross-view correspondence — without any architectural modification. This cross-view consistency further acts as a geometric regularizer, teaching the model object shape, depth, and spatial layout.\n\n* * *\n\n## Data: Embodied World Knowledge\n\n### EWK Dataset\n\n![Embodied World Knowledge Dataset](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotworld/figure_dataset_visual.png)\n\nThe Embodied World Knowledge (EWK) dataset is organized along four complementary axes, each targeting a distinct source of physical variation:\n\n- **Multi-Embodiment** — human hands, 7 robot arm configurations, ego vehicles, mobile agents, spanning 20+ distinct robot models\n- **Multi-Task** — atomic manipulation skills, long-horizon compositions, locomotion, dynamic/deformable interactions across 500+ action categories\n- **Multi-Scenario** — real-world first, sim-augmented: kitchens, workshops, outdoor settings, plus photorealistic simulation for downstream VLA evaluation\n- **Multi-View** — main, wrist, and synchronized multi-view streams (~1.6M of 6M embodied samples include 2–4 view concatenations)\n\nView Detailed Dataset Inventory\n\n| Dataset | Embodiment | Views | Contribution |\n| ----------------------------------------------- | -------------------------------- | ------------------------ | --------------------------------------------- |\n| Manipulation (~5.9M samples) | | | |\n| EgoHOD, EPIC-Kitchens, Egocentric-10k | Human hands | Egocentric | Dexterity & coordination prior |\n| Bridge V2, RH20T, Droid | Single-arm grippers | 3rd-person + wrist | Interaction primitives |\n| Robomind, RoboCoin | Single/dual-arm, humanoids | Ego + 3rd-person | Cross-embodiment generalization |\n| Agibot-World, Galaxea | Single-arm (gripper + dexterous) | Synced ego + wrist + 3rd | Temporal & multi-view consistency |\n| Qwen-Aloha (internal) | Dual-arm grippers | Head + dual wrist | Multi-view grasping prior |\n| ActionNet, OpenLoong | Dexterous hands | Wrist + 3rd-person | Fine-grained dexterity |\n| Autonomous Driving (~200K samples) | | | |\n| Waymo, NVIDIA PhysicalAI-AD, Bench2Drive, Sekai | Ego vehicle | Surround-view | Large-scale ego-motion & multi-agent dynamics |\n| Indoor Navigation (6K+ episodes) | | | |\n| VLNVerse | Mobile agent | Egocentric | Room-scale spatial reasoning |\n| Human-to-Robot Transfer | | | |\n| Scene2Robot (synthesized) | 14 robot morphologies | Multi-view | Cross-embodiment video editing |\n\n### Action-Language Mapping\n\nThe central challenge in building a universal world model is representational heterogeneity: manipulation uses joint angles, driving uses steering commands, navigation uses heading vectors — each requiring a separate model. Our **action-language mapping framework** resolves this by projecting all action signals onto a shared natural language space, so that videos from a Franka gripper, an autonomous vehicle, and a navigation agent all become instances of the same language-conditioned video generation task.\n\nA **hierarchical five-layer annotation pipeline** ensures caption quality and precision:\n\n1\n\n**Task Goal** High-level intent — what should change between states\n\n2\n\n**Action Detail** Spatio-temporal trajectories with explicit viewpoint declaration\n\n3\n\n**Physical Feedback** Observable consequences on the environment\n\n4\n\n**Comprehensive Caption** Full description for precise prediction\n\n5\n\n**Concise Caption** Essential elements for brief task-level commands\n\nDuring training, comprehensive and concise descriptions are sampled with equal probability, so the model handles both detailed trajectory specifications and brief task-level commands.\n\n* * *\n\n## Training\n\nTraining follows a **general-to-expert progressive curriculum**:\n\n| Stage | Phase | Data Mix | Objective |\n| --------------------------------- | --------------------------------- | --------------------------------- | -------------------------------- |\n| **Pretraining** | T2I / T2V / TI2V joint | General data | Build foundational visual priors |\n| Human interaction | Ego4D, EPIC-Kitchen, etc. | Grasping & tool-use priors | |\n| **SFT** | Phase 1: Single-view manipulation | Embodied + general joint training | Core manipulation physics |\n| Phase 2: Multi-view expansion | Broaden viewpoint coverage | | |\n| Phase 3: Multi-view concatenation | Cross-view geometric consistency | | |\n| Phase 4: Complex cross-domain | Long-horizon & cross-scenario | | |\n\nPretraining on general data and human interaction videos (Ego4D, EPIC-Kitchen) builds broad visual priors — the T2I task specifically anchors object geometry that transfers to video generation through the shared backbone. SFT then progressively deepens embodied expertise across four phases while keeping general data in every batch, ensuring both capabilities advance together rather than trade off.\n\n* * *\n\n## Demos\n\n### Fine-Grained Language Grounding\n\nGiven identical initial frames, the model produces qualitatively distinct videos when a single keyword differs. It also handles complex, multi-step instructions requiring long-horizon reasoning.\n\nContrastive Instruction Following:\n\nL: Pick up the red strawberry\\\nR: Pick up the yellow potato\n\nL: Place pen on the wooden tray\\\nR: Place pen on the white paper\n\nL: Hand glue to the person\\\nR: Place glue into the penholder\n\nComplex Multi-Step Instructions:\n\nSequentially pick up the red and yellow bell peppers, place them on the table from left to right\n\nGrab the yellow-and-blue stacked block, position it above the green-and-blue stacked block\n\n### Cross-Domain Generalization\n\n(A) Cross-Embodiment:\n\n(B) Cross-Task × Cross-Environment:\n\n(C) Multi-View Consistency:\n\n(D) Zero-Shot Robustness:\n\nOurs\n\nLVP\n\nCosmos2.5-14B\n\n### Human-to-Robot Transfer\n\nThe Scene2Robot mechanism preserves task intent from a human demonstration (left) while adapting motion to embodiment-specific kinematic constraints (right).\n\n### Beyond Manipulation: Driving and Navigation\n\nThe learned world model generalizes beyond robot manipulation to broader mobility scenarios.\n\n**Autonomous Driving.** Generated driving episodes from Bench2Drive, NVIDIA PhysicalAI-AD, Sekai, and Waymo demonstrate coherent scene dynamics and vehicle behaviors.\n\n**Indoor Navigation.** Egocentric navigation episodes from VLNVerse show the model’s ability to simulate first-person movement through complex indoor environments.\n\n> From tabletop manipulation to autonomous driving and indoor navigation — Qwen-RobotWorld demonstrates that a unified world model can generalize beyond a single morphology or scenario family.\n\n* * *\n\n## Performance\n\nWe evaluate against general video generation models (Sora2, Veo3, Wan2.6, Kling, LTX-2) and embodied world models (Cosmos, LVP, GigaWorld, Vidar, Wow) across four benchmarks.\n\n![Benchmark Radar Overview](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotworld/radar_figure.png)\n\nEWMBench\n\n4.60\n\nEmbodied motion fidelity\n\nDreamGen\n\n4.952\n\nInstruction following & physics alignment\n\nWorldModelBench\n\n8.99\n\nPhysical reasoning & instruction following\n\nPBench\n\n0.804\n\nPhysical behavior evaluation\n\nEWMBench\n\n| Type | Model | SceneC | HSD | Dyn | nDTW | Diversity | BLEU | CLIP | Logics | Overall |\n| --------- | -------- | --------- | --------- | --------- | --------- | --------- | ----- | ----- | --------- | -------- |\n| General | Veo3 | 0.842 | 0.213 | 0.193 | 0.161 | 0.022 | 0.214 | 0.897 | 0.947 | 3.49 |\n| Wan2.6 | 0.671 | 0.203 | 0.090 | 0.172 | 0.050 | 0.162 | 0.874 | 1.000 | 3.22 | |\n| Kling | 0.821 | 0.327 | 0.182 | 0.342 | 0.017 | 0.259 | 0.901 | 1.000 | 3.85 | |\n| LTX-2 | 0.785 | 0.208 | 0.128 | 0.244 | 0.012 | 0.143 | 0.887 | 0.500 | 2.91 | |\n| Sora2 | 0.853 | 0.281 | 0.349 | 0.275 | 0.031 | 0.247 | 0.910 | 0.947 | 3.89 | |\n| Embodied | Cosmos | 0.796 | 0.250 | 0.205 | 0.253 | 0.080 | 0.123 | 0.846 | 0.733 | 3.29 |\n| GigaWorld | 0.871 | 0.305 | 0.085 | 0.278 | 0.028 | 0.205 | 0.887 | 0.900 | 3.56 | |\n| LVP | 0.880 | 0.425 | 0.043 | 0.623 | 0.009 | 0.218 | 0.900 | 0.952 | 4.05 | |\n| Vidar | 0.734 | 0.188 | 0.152 | 0.177 | 0.065 | 0.161 | 0.882 | 0.941 | 3.30 | |\n| Wow | 0.887 | 0.249 | 0.053 | 0.257 | 0.027 | 0.193 | 0.900 | 0.952 | 3.52 | |\n| | **Ours** | **0.914** | **0.566** | **0.343** | **0.671** | 0.011 | 0.208 | 0.883 | **1.000** | **4.60** |\n\nStrong motion fidelity (HSD **0.566**), high scene consistency (**0.914**), and perfect logic constraint satisfaction.\n\nDreamGen Bench\n\n| Model | GR1-Env PA | GR1-Env IF | GR1-Object PA | GR1-Object IF | GR1-Behavior PA | GR1-Behavior IF | Total |\n| ---------- | ---------- | ---------- | ------------- | ------------- | --------------- | --------------- | --------- |\n| Cosmos-sft | 0.709 | 0.655 | 0.775 | 0.720 | 0.649 | 0.621 | 4.129 |\n| LVP | 0.810 | 0.772 | 0.745 | 0.829 | 0.713 | 0.889 | 4.758 |\n| Vidar | 0.445 | 0.647 | 0.478 | 0.726 | 0.394 | 0.651 | 3.341 |\n| GigaWorld | 0.621 | 0.933 | 0.500 | 0.852 | 0.426 | 0.884 | 4.216 |\n| Wow | 0.793 | 0.826 | 0.755 | 0.849 | 0.809 | 0.696 | 4.728 |\n| **Ours** | **0.828** | 0.793 | **0.840** | **0.878** | 0.781 | 0.832 | **4.952** |\n\nStrong object-level compositional generalization (GR1-Object IF: **0.878**), with consistent physics alignment across all subsets.\n\nWorldModelBench\n\n| Type | Model | Instr. (0-3) | Frame | Temp | Newton | Mass | Fluid | Penetr. | Grav. | Phys. | Total |\n| --------- | -------- | ------------ | ----- | ---- | -------- | -------- | -------- | ------- | -------- | -------- | -------- |\n| General | Veo3 | 2.52 | 0.98 | 0.95 | 1.00 | 0.89 | 0.99 | 0.91 | 1.00 | 4.80 | 9.25 |\n| Wan2.6 | 2.50 | 0.99 | 0.95 | 1.00 | 0.89 | 0.99 | 0.94 | 1.00 | 4.83 | 9.27 | |\n| Sora2 | 2.21 | 0.96 | 0.93 | 1.00 | 0.91 | 0.99 | 0.95 | 1.00 | 4.84 | 8.93 | |\n| Kling | 1.59 | 0.97 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 5.00 | 8.55 | |\n| LTX-2 | 1.97 | 0.69 | 0.62 | 0.99 | 0.60 | 1.00 | 0.73 | 1.00 | 4.32 | 7.61 | |\n| Embodied | Cosmos | 2.14 | 1.00 | 0.94 | 1.00 | 0.92 | 1.00 | 0.94 | 1.00 | 4.86 | 8.94 |\n| LVP | 2.01 | 0.89 | 0.91 | 1.00 | 0.93 | 0.99 | 0.95 | 1.00 | 4.87 | 8.67 | |\n| GigaWorld | 2.13 | 0.59 | 0.46 | 1.00 | 0.48 | 0.99 | 0.69 | 0.98 | 4.13 | 7.31 | |\n| Vidar | 1.62 | 0.54 | 0.45 | 1.00 | 0.56 | 1.00 | 0.85 | 1.00 | 4.40 | 7.01 | |\n| Wow | 2.05 | 0.76 | 0.65 | 1.00 | 0.65 | 0.99 | 0.81 | 1.00 | 4.45 | 7.91 | |\n| | **Ours** | **2.33** | 0.87 | 0.85 | **1.00** | **1.00** | **1.00** | 0.94 | **1.00** | **4.94** | **8.99** |\n\n**Perfect physics adherence (1.00)** across Newton's laws, mass conservation, fluid dynamics, and gravity, with strong instruction following (2.33/3.0).\n\nPBench\n\n| Type | Model | I2V-Bg | I2V-S | Aes | Img | Bg-Con | Mot | Sub-Con | O-Con | Quality | Domain | Overall |\n| --------- | -------- | ------ | ----- | ----- | ----- | --------- | --------- | ------- | ----- | ------- | --------- | --------- |\n| General | Veo3 | 0.975 | 0.980 | 0.526 | 0.698 | 0.938 | 0.994 | 0.927 | 0.128 | 0.771 | 0.882 | **0.827** |\n| Wan2.6 | 0.856 | 0.843 | 0.514 | 0.719 | 0.906 | 0.978 | 0.843 | 0.136 | 0.724 | 0.832 | 0.778 | |\n| Sora2 | 0.981 | 0.973 | 0.487 | 0.672 | 0.961 | 0.994 | 0.954 | 0.129 | 0.769 | 0.841 | 0.805 | |\n| Kling | 0.982 | 0.979 | 0.521 | 0.699 | 0.920 | 0.990 | 0.927 | 0.124 | 0.768 | 0.874 | 0.821 | |\n| LTX-2 | 0.948 | 0.955 | 0.506 | 0.622 | 0.932 | 0.986 | 0.904 | 0.118 | 0.746 | 0.845 | 0.796 | |\n| Embodied | LVP | 0.979 | 0.981 | 0.515 | 0.679 | 0.954 | 0.991 | 0.962 | 0.116 | 0.772 | 0.812 | 0.792 |\n| GigaWorld | 0.957 | 0.944 | 0.495 | 0.641 | 0.925 | 0.984 | 0.892 | 0.128 | 0.746 | 0.841 | 0.794 | |\n| Wow | 0.967 | 0.957 | 0.517 | 0.689 | 0.941 | 0.980 | 0.929 | 0.111 | 0.761 | 0.786 | 0.774 | |\n| Vidar | 0.935 | 0.922 | 0.501 | 0.573 | 0.912 | 0.982 | 0.863 | 0.120 | 0.726 | 0.810 | 0.768 | |\n| Cosmos | 0.974 | 0.973 | 0.470 | 0.663 | 0.940 | 0.989 | 0.931 | 0.160 | 0.763 | 0.840 | 0.802 | |\n| | **Ours** | 0.956 | 0.943 | 0.455 | 0.649 | **0.956** | **0.990** | 0.933 | 0.124 | 0.751 | **0.857** | **0.804** |\n\nStrong domain understanding (**0.857**) and motion smoothness (**0.990**), reflecting consistent temporal coherence across physical scenarios.\n\n[← Back to Qwen-Robot Suite](https://qwen.ai/blog?id=qwen-robotsuite)\n\n* * *\n\n## Citation\n\nIf you find our work helpful, feel free to give us a cite.\n\n```bibtex\n@article{qwenrobot-world,\n title={Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation},\n author={Qwen Team},\n year={2026}\n}\n```\n", "dateCreated": "2026-10-19T01:26:53.998049939+00:00", "date_created": "2026-06-16T00:00:00+00:00", "date_modified": null, "date_published": "2026-06-16T00:00:00+00:00", "description": "Embodied intelligence requires agents to perceive, reason about, and act within physical environments. World models offer a scalable path forward — but current approaches face a fundamental tension. General video generation models learn…", "external_url": "https://qwen.ai/blog?id=qwen-robotworld", "git": { "commit": "ebb0941c5c06a2ab2377342d607535262a3e969e", "path": "items/qwen-ai/2026/06/2026-06-16-qwen-robotworld-boundless-worlds-for-embodied-agents.md", "permalink": "https://github.com/aymericbeaumet/aggr-instance/blob/ebb0941c5c06a2ab2377342d607535262a3e969e/items/qwen-ai/2026/06/2026-06-16-qwen-robotworld-boundless-worlds-for-embodied-agents.md" }, "id": "urn:aggr:item:e794adae88220fcca61b7bf15eff88a39cb72953", "isBasedOn": { "@id": "https://qwen.ai/blog?id=qwen-robotworld" }, "mainEntity": { "@id": "https://qwen.ai/blog?id=qwen-robotworld", "@type": "CreativeWork", "archivedAt": { "@id": "https://aggr.aymericbeaumet.com/items/qwen-ai/2026-06-16-qwen-robotworld-boundless-worlds-for-embodied-agents/#webpage" }, "author": [ { "@type": "Person", "name": "QwenTeam" } ], "datePublished": "2026-06-16T00:00:00+00:00", "headline": "Qwen-RobotWorld: Boundless Worlds for Embodied Agents", "image": { "@type": "ImageObject", "caption": null, "height": 90, "url": "https://img.alicdn.com/imgextra/i2/O1CN01tVjHlI24sKFloAjuy_!!6000000007446-2-tps-1590-954.png", "width": 160 }, "timeRequired": "PT9M", "url": "https://qwen.ai/blog?id=qwen-robotworld", "wordCount": 1933 }, "name": "Archived snapshot: Qwen-RobotWorld: Boundless Worlds for Embodied Agents", "preview": { "alt": null, "color": null, "height": 90, "placeholder": { "data_url": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=", "hash": "" }, "url": "https://img.alicdn.com/imgextra/i2/O1CN01tVjHlI24sKFloAjuy_!!6000000007446-2-tps-1590-954.png", "width": 160 }, "reading_minutes": 9, "replicated_at": null, "source": "qwen-ai", "source_name": "Qwen", "summary": null, "tags": [ "release" ], "temporalCoverage": "2026-10-19T01:26:53.998049939+00:00", "title": "Qwen-RobotWorld: Boundless Worlds for Embodied Agents", "url": "https://aggr.aymericbeaumet.com/items/qwen-ai/2026-06-16-qwen-robotworld-boundless-worlds-for-embodied-agents/", "word_count": 1933 }