Qwen Robot Suite: Unified Foundation Models for Navigation, Manipulation, and World Modeling

TL;DR

Qwen announced the Qwen‑Robot Suite, a trio of foundation models—Qwen‑RobotNav, Qwen‑RobotManip, and Qwen‑RobotWorld—that translate language‑level intelligence into concrete physical actions, enabling a single agentic system to navigate, manipulate, and predict world dynamics across dozens of robot embodiments.

Overview of the Qwen‑Robot Suite

The Qwen‑Robot Suite is designed to close the long‑standing gap between vision‑language understanding and embodied control. It does so by providing three specialized foundation models, each aligning natural‑language instructions with a distinct action domain:

  • Qwen‑RobotNav – converts language into mobility commands for navigation and driving.
  • Qwen‑RobotManip – maps language to manipulation actions across heterogeneous robot arms.
  • Qwen‑RobotWorld – predicts future video frames from language‑conditioned actions, effectively learning a world dynamics model.

All three expose language‑first interfaces, allowing higher‑level Qwen models (e.g., Qwen‑3.7‑Plus) to invoke them as reusable tools within a broader agentic architecture.

Qwen‑RobotNav: Unified Mobility Across Five Domains

Key technical contributions

  • Built on Qwen3‑VL, the model introduces a parameterised navigation interface with two axes: task mode (instruction following, object search, tracking, autonomous driving, embodied question answering) and a four‑axis controllable observation protocol (visual token budget, temporal decay, per‑camera weighting, frame sampling).
  • Trained on 15.6 M navigation samples while co‑training on vision‑language data to preserve grounded perception.
  • Uses a two‑tier system: an upper‑level planner (Qwen‑3.7‑Plus) decomposes long‑horizon goals and dynamically switches task modes and observation settings mid‑episode.

Performance highlights

  • State‑of‑the‑art success rates on five benchmarks: 76.5 % SR on VLN‑CE RxR, 75.6 % SR on HM3Dv2 object‑goal (RGB‑only), 90.0 % tracking rate on EVT‑Bench, 91.4 % PDMS on NAVSIM, and new bests on three Embodied QA datasets.
  • Scales log‑linearly from 2 B to 8 B parameters.
  • Zero‑shot deployment on a Unitree Go2 quadruped (NVIDIA Jetson Thor, 196 ms latency) using only the built‑in low‑resolution camera, successfully following multi‑room natural‑language instructions.

Implications The controllable observation protocol enables a single model to serve as a reconfigurable navigation primitive for any agentic system, eliminating the need for task‑specific architectures and supporting long‑horizon reasoning with persistent memory.

Qwen‑RobotManip: Cross‑Embodiment Manipulation Foundation

Core alignment mechanisms

  1. Unified 80‑dimensional state‑action space shared by single‑arm, dual‑arm, dexterous‑hand, and mobile robots.
  2. Camera‑frame end‑effector delta poses that make visually similar motions numerically close, abstracting away morphological differences.
  3. In‑context policy adaptation that reads execution history as an implicit embodiment signature, allowing on‑the‑fly adaptation.

Training data

  • >38 100 hours of open‑source data, comprising:
    • 11 320 h of robot demonstrations,
    • 1 933 h of egocentric human video,
    • 24 808 h of robot demonstrations synthesized from the human video across 15 embodiments via a Human‑to‑Robot pipeline (action retargeting, hand removal, simulated rendering, depth‑guided compositing).

Benchmark results

  • LIBERO‑Plus: 91.4 % ( +7.0 over π0.5 baseline )
  • RoboTwin‑C2R Hard: 69.4 % ( +21.5 )
  • RoboCasa365 Composite‑Unseen: 14.9 % ( 3× next best )
  • EBench: 45.6 % ( +18.5 )
  • RoboTwin‑IF: 72.0 % ( +22.4 )
  • Ranked #1 on the RoboChallenge Table30 v1 Generalist Track, sweeping the top‑2 positions and leading the third place by 20 %.

Implications Unified representation is shown to be a prerequisite for scaling: only models with the UnifiedSpace + UnifiedEEF formulation exhibit clean log‑linear performance gains as data volume increases. This enables a single policy to handle diverse manipulation tasks, perform zero‑shot cross‑embodiment transfer, and recover from errors through high‑level replanning.

Qwen‑RobotWorld: Language‑Conditioned World Modeling

Design philosophy

  • Actions are expressed solely in natural language, unifying end‑effector poses, steering commands, and navigation waypoints under a single interface.
  • Trains on the Embodied World Knowledge corpus: 8.6 M video‑text pairs covering >200 M frames, spanning 20+ robot embodiments and 500+ action categories.

Model architecture

  • A 60‑layer dual‑stream MMDiT couples Qwen2.5‑VL semantic embeddings with video latents.
  • The action encoder is a full multimodal LLM (not a lightweight text encoder), allowing the model to internalize physical knowledge (rigid bodies, fluid dynamics, gravity) and generate physically plausible futures.

Performance and capabilities

  • 1st place overall on EWMBench (motion fidelity +33 % over runner‑up) and DreamGen Bench.
  • 1st open‑source on WorldModelBench (perfect adherence to Newtonian physics, mass conservation, fluid dynamics) and PBBench.
  • Generates fine‑grained, keyword‑sensitive futures (e.g., swapping “red strawberry” for “yellow potato” changes the predicted outcome accordingly).
  • Produces temporally and spatially consistent multi‑view videos, facilitating sim‑to‑real transfer and multi‑camera policy training.
  • Enables zero‑shot human‑to‑robot transfer: a single human demonstration can be turned into realistic robot executions across multiple embodiments (e.g., Kinova Gen3, KUKA iiwa, xArm7).

Implications By learning a unified world dynamics model, Qwen‑RobotWorld serves as both a synthetic data engine and a closed‑loop evaluator for other suite components, allowing manipulation, navigation, and driving tasks to reinforce each other’s physical understanding.

From Models to Agents: Closing the Loop

The suite’s language‑first APIs let general‑purpose Qwen models act as high‑level planners while delegating low‑level execution to the specialized suite models. Demonstrated applications include:

  • Open‑ended task execution: Qwen‑Omni proposes manipulation tasks via speech, evaluates outcomes in real time, and hands execution to Qwen‑RobotManip without a predefined task list.
  • Long‑horizon manipulation: A Qwen‑3.5 planner decomposes a complex tabletop cleaning instruction into atomic subtasks; Qwen‑RobotManip executes each subtask, enabling robust performance on out‑of‑distribution scenes and automatic replanning after failures.
  • Agentic navigation & embodied QA: Combining Qwen‑RobotNav with a high‑level planner yields state‑of‑the‑art results on Embodied Question Answering benchmarks (HM‑EQA, MT‑HM3D, EXPRESS‑Bench) and real‑world demos such as locating an open restroom in a building.
  • Chat2Robot demo: An experimental web interface lets users type natural‑language commands that are executed by Qwen‑RobotManip in real time (currently limited to 50 tasks from the RoboTwin‑Clean dataset).

These integrations illustrate how the suite can serve as a physical‑world toolset for any multimodal LLM, turning abstract reasoning into concrete robot actions.

Future Directions

Qwen identifies several open challenges:

  • Contact‑rich, long‑horizon tasks that require continual learning.
  • Tighter, low‑latency coupling between general‑purpose planners and physical executors.
  • richer human‑robot‑environment interaction modalities (e.g., tactile feedback, speech).

The roadmap emphasizes scaling multimodal perception, refining the language‑action alignment, and expanding the suite to cover more embodiments and action categories.

Citation

@article{qwenrobotnav,
  title   = {Qwen-RobotNav: A Scalable Navigation Model Designed for an Agentic Navigation System},
  author  = {Qwen Team},
  year    = {2026}
}
@article{qwenrobotmanip,
  title   = {Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models},
  author  = {Qwen Team},
  year    = {2026}
}
@article{qwenrobotworld,
  title   = {Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation},
  author  = {Qwen Team},
  year    = {2026}
}

Sources