Qwen-Robot Suite - A physical world basic model kit launched by Alitongyi - AiBoss

What is Qwen-Robot Suite?

Qwen-Robot Suite is a foundational model suite launched by Alibaba Tongyi for achieving physical world intelligence. It includes three basic models: Qwen-RobotNav (navigation), Qwen-RobotManip (operation), and Qwen-RobotWorld (world model). The suite bridges the visual language representation space to physical actions through a language-first interface, unifying training and inference across ontologies and scenarios. This enables the general-purpose large model to act as a high-level planner, dynamically calling underlying physical tools to complete long-term task loops.

Main functions of Qwen-Robot Suite

Unified Physical NavigationQwen-RobotNav covers five major areas: instruction following, point/target navigation, target tracking, autonomous driving, and embodied question answering, all executed by a single model.
Cross-body operation controlQwen-RobotManip is compatible with 15 robot bodies, including single-arm, dual-arm, dexterous hand, and mobile platform, enabling direct mapping from visual language to actions.
World Dynamics ForecastQwen-RobotWorld receives natural language action descriptions and predicts future states that conform to physical laws in operation, driving, and navigation scenarios.
Intelligent Agent ToolkitAll three models provide a language-first interface, which can be dynamically invoked and combined by the general Qwen model as a tool for the physical world.
Zero-sample real deploymentSupports zero-sample deployment on real robots such as Unitree Go2, without requiring scenario-specific fine-tuning.

Technical Principles of Qwen-Robot Suite

Qwen-RobotNav: Parametric Vision Assignment and Controllable Observation ProtocolThe model parameterizes the vision assignment strategy itself, selecting navigation behavior (command following, target search, tracking, autonomous driving) through task modes, and exposing four-axis inference parameters—visual token budget, time decay, per-camera weights, and frame sampling mode—as adjustable interfaces. The model is trained on 15.6 million samples combined with visual and language data, allowing the same set of weights to adapt to different navigation needs without modifying the architecture. The upper-level planner dynamically switches between task modes and context strategies, repeatedly calling the model as a reconfigurable navigation primitive, and employing a two-layer memory mechanism to achieve long-range inference.
Qwen-RobotManip: Unifying Cross-Ontology Alignment and Scalable TrainingUsing Qwen3.5-4B VL as the backbone and combining it with a flow-matching DiT action head, this study resolves heterogeneous ontology conflicts through three mechanisms. A unified 80-dimensional state-action representation is shared across diverse ontology types; incremental pose positioning of the end effector in the camera coordinate system ensures that visually similar movements are numerically similar, masking morphological differences; finally, context-adaptive policy treats execution history as an implicit ontology identifier, enabling online behavior calibration during inference. In terms of data, training was completed using 11,320 hours of open-source robot data, 1,933 hours of human video, and 24,808 hours of cross-ontology data generated by the human-robot transfer synthesis pipeline, totaling over 38,100 hours of corpus.
Qwen-RobotWorld: A Language-Driven Two-Stream World ModelThe key design element is to unify all actions into a natural language interface, standardizing end-effector pose, steering commands, and navigation landmarks into a single training signal, enabling collaborative training of over 20 ontology types and over 500 action categories. The model employs a 60-layer dual-stream MMDiT, deeply coupling Qwen2.5-VL semantic representations with video latent variables. Unlike lightweight text encoders, this model uses a complete multimodal large language model as the action encoder, internalizing physical knowledge such as rigid bodies, fluids, and gravity, implicitly generating a physically believable future.

How to use Qwen-Robot Suite

Select ModuleChoose RobotNav (mobility), RobotManip (operation), or RobotWorld (prediction) as the underlying physics engine based on task requirements.
Configuration InterfaceThe visual history encoding parameters can be adjusted through a four-axis controllable observation protocol to adapt to different input configurations of single or multiple cameras.
Access plannerThe upper-level general-purpose intelligent agent (such as Qwen-3.5) decomposes long-term goals into sub-tasks and dynamically calls the corresponding models through language commands.
Execution closed loopThe underlying model receives language instructions and outputs physical actions or future frames to drive the execution of real robots or simulation environments.
Memory ManagementIt leverages cross-fragment memory and evidence notebooks to compress context, supporting multi-step navigation, environmental forensics, and failure recovery.

The core advantages of Qwen-Robot Suite

Language-first unified interfaceAll three models provide natural language action interfaces, which can be seamlessly called as tools by general LLMs without the need to design proprietary protocols for different robot forms.
Cross-ontology strong generalizationA unified 80-dimensional representation aligned with camera coordinate system motion enables the model to migrate across 15 platforms under zero-sample conditions, achieving twice the previous best performance in real-world environments.
Training with fully open-source dataRobotManip completed over 38,000 hours of training using only open-source data, significantly reducing data barriers through a human-machine transfer synthesis pipeline, proving that alignment is a prerequisite for scalability.
Physical consistency generationThe world model uses a complete VLM as its motion encoder, internalizing Newton's laws, conservation of mass, and fluid dynamics, and ranks first among open-source models on physical inference benchmarks such as WorldModelBench.

Qwen-Robot Suite project address

Model weights, technical blog, and independent project page may not be fully public yet.

Comparison of Qwen-Robot Suite with similar products

Comparison Dimensions	Qwen-Robot Suite	π0.5 (Physical Intelligence)
Data Strategy	Training on fully open-source data (>38K hours), including a human-machine transfer synthesis pipeline.	Relying on large-scale real robot teleoperation data collection
Body compatibility	A unified 80-dimensional representation, spanning 15 platforms and over 20 ontologies.	Primarily optimized for specific ontology types, with limited cross-morphological generalization.
Action Interface	Language-first approach, unifying operation, navigation, and driving tasks using natural language.	Action tokens are the primary component, with language as a secondary factor.
Generalization ability	Zero-sample cross-ontology transfer, real-world deployment in open environments	Strongly distributed performance, but OOD generalization and cross-ontology transfer are relatively limited.
System Architecture	A three-model suite combined with an agent framework provides hierarchical support for long-range planning.	A single VLA model executes end-to-end, while long-term tasks rely on external planning.

Application scenarios of Qwen-Robot Suite

Home service robotsIt can perform long-term operations such as desktop cleaning, clothing organization, and cooking assistance, and has the ability to break down tasks and recover from failures.
Warehousing and LogisticsThe same navigation model handles target search, shelf inventory and AGV path tracking within the warehouse, enabling zero-sample deployment across different models.
Autonomous driving and embodied navigationThe unified model covers urban driving, indoor command following, and open environment exploration, and supports evidence-based question answering.
Simulation data synthesisGenerate physically plausible future frames using a world model as a low-cost, large-scale training data engine.
Human-robot collaborative industrial armIt enables skill transfer across 15 industrial/service robot platforms, reducing reprogramming costs during production line changeovers.