Qwen-Robot Suite: A Foundation Model Suite for Physical World Intelligence

The Qwen-Robot Suite bridges the gap between vision-language understanding and physical action by providing a set of foundation models designed for mobility, interaction, and world dynamics. While multimodal foundation models can plan tasks in language, they often lack the ability to produce the specific motor commands required for execution. The Qwen-Robot Suite addresses this alignment challenge by aligning language instructions with physical action signals across diverse robotic embodiments.

Qwen-RobotNav: Unified Physical Mobility

Qwen-RobotNav is a foundation model for mobility that unifies five navigation task families—including instruction following, object search, target tracking, autonomous driving, and embodied question answering (EQA)—under a single set of weights. Built on Qwen3-VL, it uses a parameterized navigation interface that allows agentic systems to dynamically switch task modes and observation parameters (such as token budget and temporal decay) during inference.

Key technical achievements include:

  • Multi-Domain Performance: The model achieves state-of-the-art (SOTA) results across five domains, including 76.5% SR on VLN-CE RxR and 91.4 PDMS on NAVSIM.
  • Controllable Observation Protocol: It enables inference-time configuration of visual history encoding without requiring architectural changes or retraining.
  • Agentic Integration: When paired with an upper-level planner like Qwen3.7-Plus, it improves performance on EXPRESS-Bench by 15.4%, reducing navigation steps by 77%.
  • Zero-Shot Generalization: The model was successfully deployed zero-shot on a Unitree Go2 quadruped robot using only its built-in low-resolution camera in unseen environments.

Qwen-RobotManip: Scaling Robotic Interaction

Qwen-RobotManip focuses on the challenge of heterogeneous robot embodiments. It enables cross-embodiment training by introducing a unified 80-dimensional state-action representation and camera-frame end-effector delta poses, which make visually similar motions numerically proximate across different robot arms.

Key technical achievements include:

  • Large-Scale Data Training: The model was trained on over 38,100 hours of data, including 11,320 hours of open-source robot data and 24,808 hours of robot demonstrations synthesized from human video via a Human-to-Robot synthesis pipeline.
  • OOD Generalization: The model shows strong out-of-distribution (OOD) generalization, significantly outperforming previous models on benchmarks like LIBERO-Plus (91.4%) and RoboTwin-C2R Hard (69.4%).
  • Real-World Validation: It ranked #1 on the RoboChallenge Table30 v1 generalist track with a 45% success rate.
  • Scaling Law: The team found that unified cross-embodiment representations are a prerequisite for log-linear data scaling; without this alignment, adding more data does not lead to consistent performance gains.

Qwen-RobotWorld: Predictive World Dynamics

Qwen-RobotWorld is a world model that predicts the next state of the physical world given a current observation and a natural-language action. By expressing all actions in natural language, the model can co-train across 20+ robot embodiments and 500+ action categories.

Key technical achievements include:

  • Unified Action Interface: It utilizes a dual-stream MMDiT and a Qwen2.5-VL action encoder to parse complex compositional instructions into physically grounded futures.
  • Performance Benchmarks: It ranks 1st overall on EWMBench and DreamGen Bench, and is the top open-source model on WorldModelBench and PBBench.
  • Capabilities: The model demonstrates fine-grained language grounding, multi-view consistent generation, and the ability to transfer human demonstrations to robot execution across eight different embodiments (e.g., xArm7, Franka Panda, Sawyer).

From Models to Agents: Closing the Loop

Because the Qwen-Robot Suite models use language-first interfaces, they can be composed as tools by general-purpose Qwen VLMs. This is implemented through the Qwen-RobotClaw agent harness, which allows a high-level planner (such as Qwen-3.5) to decompose complex instructions into atomic subtasks for the VLA (Vision-Language-Action) models to execute.

Practical applications of this agentic approach include:

  • Open-Ended Task Execution: Qwen-Omni can propose and evaluate manipulation tasks in real-time without a pre-defined list.
  • Long-Horizon Manipulation: By decomposing tasks (e.g., table cleaning) into subtasks, the system improves robustness to OOD scenes and failures, allowing for recovery and replanning.
  • Embodied QA: Combining the agent system with Qwen-RobotNav allows for complex exploration tasks, such as finding an open restroom in a real building by interpreting visual cues like "Cleaning in Progress" signs.

Community Insights and Perspectives

Discussion among technical observers suggests that the integration of world models with robotic embodiment provides a path toward genuine artificial intelligence by creating a continuous feedback loop of action and result.

"In my opinion, training through embodiment and constructing an internal world model makes it possible to do genuine reasoning about how objects behave in the physical world... Having a world model in the context of robotics is the most likely path towards creating a genuine artificial intelligence."

Other observers noted the strategic importance of this suite for manufacturing and humanoid robotics, suggesting that the scale of the Qwen ecosystem could lead to rapid mass production of integrated robotic systems.

Sources