認識 Qwen-RobotSuite：三個用於 VLA 操作、影片世界建模與導航的具身 AI 模型

The Qwen team has released three embodied AI models, grouped as Qwen-Robot-Suite. The three are Qwen-RobotManip, Qwen-RobotWorld, and Qwen-RobotNav. Each is built on a Qwen vision-language backbone and targets a different robotics problem. Qwen-RobotManip is a Vision-Language-Action model for manipulation, built on Qwen3.5-4B. Qwen-RobotWorld is a language-conditioned video world model with a 60-layer MMDiT and a frozen Qwen2.5-VL encoder. Qwen-RobotNav is a navigation model built on Qwen3-VL, available at 2B, 4B, and 8B sizes. Qwen-Robot-Suite Qwen-Robot-Suite is not a single model. It is a suite of three independent foundation models. Two of them, RobotManip and RobotNav, ship with public GitHub repositories. Robotics data is fragmented across hardware and tasks. Different robots use incompatible observation and action formats. A policy trained on one arm rarely transfers to another. The three research reports address this fragmentation in different ways. RobotManip aligns action representations so manipulation data scales. RobotWorld uses language as a unified action interface for video prediction. RobotNav exposes a controllable observation interface for navigation tasks. Here is the core split between the three releases: ModelProblemBackboneOutputQwen-RobotManipRobotic manipulationQwen3.5-4B (Qwen-VL)Continuous robot actionsQwen-RobotWorldEmbodied world modelingFrozen Qwen2.5-VLPredicted future videoQwen-RobotNavMobile navigationQwen3-VL (2B/4B/8B)Waypoint trajectories Qwen-RobotManip: Alignment Unlocks Scale for Manipulation Qwen-RobotManip is a Vision-Language-Action (VLA) foundation model. It is built on Qwen-VL and predicts continuous robot actions. A VLA model takes camera views and a language instruction. It then outputs low-level robot actions. The challenge is that manipulation data is heterogeneous by nature. Different robots record states and actions in incompatible formats. When demonstrations arrive with mismatched representations, scaling data produces interference. RobotManip solves this with a unified alignment framework. The Unified Alignment Framework The framework has three complementary mechanisms. First is a canonical state-action representation. It is an 80-dimensional vector with per-dimension binary masking. This vector holds two 29-dimensional per-arm blocks plus 22 reserved dimensions. Each block stores joint positions, end-effector pose, gripper state, and dexterous hand joints. Robots populate only the dimensions they have. Second is a camera-frame delta pose parameterization. End-effector actions are expressed as deltas in the camera frame. This makes visually similar motions numerically proximate across embodiments. Third is an in-context policy adaptation mechanism. It reads recent execution history as an implicit embodiment identifier. The policy adjusts behavior at deployment time without parameter updates. A dual-stream co-training strategy runs alongside this. It jointly optimizes manipulation data and a vision-language stream. This prevents the backbone’s perception and reasoning from eroding. The Data Engine RobotManip assembles roughly 38,100 hours of manipulation data. It uses only open-source datasets and human videos. No proprietary data collection was used. A human-to-robot synthesis pipeline produces most of this scale. It converts egocentric hand demonstrations into robot trajectories. The pipeline renders across 15 robot platforms. This synthesis alone yields about 24,808 hours of demonstrations. The egocentric source data is about 1,933 hours. Open-source robot datasets contribute over 11,000 hours. The pipeline separates action alignment from visual alignment. Action alignment retargets hand keypoints to gripper poses. Visual alignment uses SAM3 masking, ProPainter inpainting, and MuJoCo inverse kinematics. A five-stage curation pipeline then filters the combined corpus. It catches sudden changes, temporal misalignment, and extreme values. One check found 81% of episodes in a subset failed state-action alignment. Benchmark Results The research report argues standard benchmarks fail to measure generalization. Models without robot pretraining match pretrained ones on in-distribution tests. RobotManip therefore focuses on out-of-distribution (OOD) settings. Benchmark (OOD)Prev. SOTA (π0.5)Qwen-RobotManipLIBERO-Plus84.491.4RoboTwin-C2R Hard47.969.4EBench27.145.6RoboCasa36516.935.9RoboTwin-IF49.672.2 The largest reported gap is on cross-embodiment transfer. RobotManip reaches 23.9% using camera-frame EEF actions. That is 3.2× the 7.5% achieved by π0.5. The model also ranks 1st on the RoboChallenge Table30-v1 generalist track. It scores a 20% relative improvement over the prior best. Real-robot validation covers AgileX ALOHA, Franka, UR, and ARX platforms. (function(){ var frame=document.getElementById("qwen-robotmanip-canonical-vector-frame"); window.addEventListener("message",function(e){ if(e&&e.data&&e.data.type==="rmcv-resize"&&typeof e.data.height==="number"){ frame.style.height=Math.max(200,Math.min(4000,e.data.height))+"px"; } }); })(); Qwen-RobotWorld: Language as a Universal Action Interface Qwen-RobotWorld is a language-conditioned video world model. It predicts future visual trajectories from a current observation. Natural language serves as the unified action interface. A world model learns environment dynamics. Given a current state and an action, it predicts the next state. RobotWorld represents states as video frames and actions as text. This is important because language is embodiment-agnostic. One instruction encodes the action sequence, goal, and constraints. It works across a Franka gripper, an Aloha dual-arm system, or a humanoid. The Double-Stream MMDiT Architecture The model uses a 60-layer double-stream Multimodal Diffusion Transformer. An understanding stream processes a frozen Qwen2.5-VL encoder’s features. A generation stream processes video-VAE latents. The two streams interact via joint attention at every layer. Using an MLLM as the action encoder gives two advantages. It parses compositional instructions and constrains physically plausible transitions. The MMDiT has 20B parameters. The VAE adopts the Wan-VAE architecture. The context length supports up to 48,360 video tokens. A Scene2Robot mechanism reuses this backbone for cross-embodiment synthesis. It processes scene, robot reference, and generation segments together. This enables human-to-robot video transfer without robot-specific prompting. The Embodied World Knowledge Dataset Training uses the Embodied World Knowledge (EWK) dataset. It contains roughly 8.6M video-text pairs. That spans over 200M observation frames. The corpus covers four embodied domains plus general video. Manipulation provides about 5.9M samples across 20+ morphologies. Driving, navigation, and human-to-robot transfer fill out the rest. An action-language mapping framework standardizes everything. It converts 20+ embodiment types and 500+ action categories into language. A hierarchical five-layer annotation pipeline produces the captions. Benchmark Results RobotWorld was evaluated on four established benchmarks. It ranks 1st overall on two of them: BenchmarkResultRankingEWMBench4.601st overallDreamGen Bench4.9521st overallWorldModelBench8.991st open-source (3rd overall)PBench0.8041st open-source On EWMBench it leads motion fidelity with an HSD of 0.566. That is a 33% gain over the runner-up. Scene consistency reaches 0.914. On WorldModelBench it scores 1.00 on four physics-adherence categories. These are Newton’s laws, mass conservation, fluid dynamics, and gravity. Penetration scores 0.94, and instruction following scores 2.33 out of 3.0. (function(){ var frame=document.getElementById("qwen-robotworld-language-interface-frame"); window.addEventListener("message",function(e){ if(e&&e.data&&e.data.type==="rww-resize"&&typeof e.data.height==="number"){ frame.style.height=Math.max(200,Math.min(4000,e.data.height))+"px"; } }); })(); Qwen-RobotNav: A Con