--- title: 'Qwen-RobotNav: A Scalable Navigation Model Designed for an Agentic Navigation System' link: https://qwen.ai/blog?id=qwen-robotnav source: qwen-ai published: 2026-06-16T00:00:00Z first_seen: 2026-10-19T01:26:53.998049939Z authors: - QwenTeam labels: - release content: feed html: 2026-06-16-qwen-robotnav-a-scalable-navigation-model-designed-for-an.html remote_preview: url: https://img.alicdn.com/imgextra/i4/O1CN01ttKuLR1VWn4v7hWsE_!!6000000002661-2-tps-1590-954.png --- [GitHub](https://github.com/QwenLM/Qwen-RobotNav) [Paper](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/papers/Qwen_RobotNav.pdf) ![Qwen-RobotNav banner](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/suite/singlenav.png) Agentic navigation systems require a base navigation model with a configurable navigation context protocol: instruction following, object search, target tracking, and autonomous driving share the same perception-planning backbone yet demand fundamentally different context strategies for consuming the visual stream. Like the Model Context Protocol for LLM tool use, a navigation model needs a standardised interface to manage these diverse context requirements at inference time. ![Qwen-RobotNav Overview](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotnav/fig_teaser.png) We present **Qwen-RobotNav**, a scalable navigation model built on Qwen3-VL that addresses this through a **parameterised interface** with two complementary dimensions: task modes that select the navigation behaviour, and controllable observation parameters (token budget, temporal decay, per-camera weights) that govern how visual history is encoded. Trained on **15.6 million samples** with training-time randomization over all parameters, Qwen-RobotNav generalizes to any inference-time configuration without architectural modification, unifying five task families under a single set of weights and serving as a natural building block for agentic systems. **Highlights:** 5 Domains\ 8 SOTAs One Model Unifies VLN, ObjNav,\ Tracking, Driving & EQA Context\ \= Interface 4-Axis Observation Protocol\ Zero Architecture Change Agentic Navigation as a Tool Call\ \+ Two-Level Memory\ Raises the new bar on EQA Generalization In-the-Wild Single-Camera\ Deployment in Unseen Environments - **Unified Multi-Domains Navigation:** a single model with one set of weights achieves state-of-the-art across 5 navigation domains: 76.5% SR on VLN-CE RxR, 75.6% SR on HM3Dv2 object-goal (RGB only, surpassing depth-based methods), 90.0% tracking rate on EVT-Bench, 91.4 PDMS on NAVSIM, and new bests on 3 EQA benchmarks, with consistent scaling from 2B to 8B parameters. - **Controllable Observation Protocol:** four axes (visual token budget, temporal decay, per-camera weighting, frame sample mode) are exposed as inference-time parameters and randomized per sample at training time, enabling any inference-time configuration without retraining or architectural modification. - **Agentic Navigation System:** designed as a reconfigurable navigation primitive within a two-tier system, where an upper-level planner (Qwen3.7-Plus) decomposes long-horizon goals and dispatches configurable navigation calls while maintaining two-level memory, achieving +15.4% on EXPRESS-Bench with 77% fewer navigation steps over prior best. - **In-the-Wild Generalization:** deployed zero-shot on a Unitree Go2 quadruped with its single low-resolution build-in camera, demonstrating strong generalization to in-the-wild environments and unconstrained natural-language instructions without any environment-specific fine-tuning. * * * ## A Navigation Model with Controllable Context Mobile navigation spans tasks with wildly different needs. Instruction following demands a long memory of past observations to re-reference distant landmarks. Target tracking cares almost entirely about the most recent few frames. Object search shifts mid-episode: broad history during exploration, tight recency during approach. Existing unified navigation models embed a single assumption about what to remember. Qwen-RobotNav addresses this by treating context as a first-class, externally controllable degree of freedom. The model provides four control axes that an agent can tune per call: - **Visual token budget**: total tokens across all cameras and timesteps - **Temporal decay**: how strongly recent frames are favored over older ones - **Camera weights**: per-camera importance (the forward camera matters more than the rear) - **Frame sample mode**: random for global history coverage, or latest for a tight recency window At training time, all these parameters are randomized per sample. The model never sees a fixed configuration, so it generalizes to any setting at inference time. ![Qwen-RobotNav Architecture](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotnav/fig_pipeline.png) **Architecture.** Qwen-RobotNav inherits the Qwen3-VL backbone and adds a lightweight 4-layer MLP action head that outputs 8 waypoints, each with position and heading. Camera identity and temporal order are communicated entirely through natural-language tags interleaved with visual tokens: ``` Time step 0 Front View Front Right View ... Time step 1 Front View ... ``` * * * ## Built for Agentic Navigation Qwen-RobotNav is designed as a **reconfigurable navigation primitive** within a two-tier system. An upper-level planner (Qwen3.7-Plus) decomposes long-horizon goals into sub-goals, while Qwen-RobotNav executes each navigation segment as a reactive waypoint predictor. > The planner can dynamically switch Qwen-RobotNav’s task mode and context strategy mid-episode. The two tiers communicate entirely through natural language, keeping the system modular and extensible. ![Agentic Navigation System](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotnav/fig_agents_overview.png) Each navigation call specifies three things: a sub-goal instruction, a task mode (VLN / PointNav / ObjNav / Tracking), and an observation configuration. The same model weights serve all task phases; only the call arguments change. To support long-horizon reasoning, the system keeps a two-level memory. Each navigation segment produces a compact trajectory summary. A persistent evidence notebook accumulates durable conclusions across episodes (searched regions, candidate object locations, rejected hypotheses) so the planner always works with concise, relevant context. * * * ## Training at Scale Qwen-RobotNav is trained on **15.6 million samples** across five task families, plus vision-language reasoning data that preserves the backbone’s perceptual grounding. Performance scales consistently from 2B to 8B parameters, with the most pronounced gains on long-horizon reasoning tasks. ![Training Data Distribution](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotnav/fig_data_distribution.png) ![Scaling Behavior](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotnav/fig_scaling.png) We also introduce an automated pipeline that converts **text-to-video generations into navigation trajectories** through prompt generation, video synthesis, VLM quality filtering, monocular depth estimation, and kinematic filtering, yielding 40K additional photorealistic samples without any 3D scene reconstruction. * * * ## Performance Instruction Following (VLN-CE) Development trend of VLN success rates on R2R and RxR validation-unseen splits. Traditional methods show steady early progress, while recent LLM/VLM-based approaches drive the latest performance gains, with Qwen-RobotNav achieving the highest success rates on both benchmarks. ![VLN-CE Performance Overview](https://qianwen-res.oss-accelerate.aliyuncs.com/qwenrobot/robotnav/vlnce_overview.png) Qwen-RobotNav-8B achieves **72.1% SR** on R2R and **76.5% SR** on the longer-horizon RxR benchmark, surpassing prior best methods. | Method | R2R SR↑ | R2R SPL↑ | RxR SR↑ | RxR SPL↑ | | -------------------- | -------- | -------- | -------- | -------- | | NaVILA | 54.0 | 49.0 | 49.3 | 44.0 | | NavFoM | 61.7 | 55.3 | 64.4 | 56.2 | | ABot-N0 | 66.4 | 63.9 | 69.3 | 60.0 | | OmniNav | 69.5 | 66.1 | 73.6 | 62.0 | | **Qwen-RobotNav-4B** | 69.5 | 63.6 | 75.2 | 65.0 | | **Qwen-RobotNav-8B** | **72.1** | **66.6** | **76.5** | **65.7** | Object Searching On HM3Dv2 object-goal navigation, Qwen-RobotNav-4B achieves **75.6% SR** using only RGB observations, surpassing all depth-based methods, and reaches only **1.72m** from the goal on average. | Method | SR↑ | SPL↑ | | -------------------- | -------- | -------- | | VLFM | 52.5 | 30.4 | | CogNav | 72.5 | 26.2 | | Uni-NaVid | 73.7 | **37.1** | | **Qwen-RobotNav-4B** | **75.6** | 30.6 | | **Qwen-RobotNav-8B** | 71.2 | 33.0 | Active Visual Tracking On EVT-Bench, Qwen-RobotNav achieves the highest tracking rate at **90.0%**, exceeding dedicated trackers and generalist models alike. | Method | TR↑ | CR↓ | SR↑ | | -------------------- | -------- | -------- | -------- | | TrackVLA++ | 81.0 | **2.10** | 86.0 | | NavFoM | 80.5 | — | 85.0 | | ABot-N0 | 87.6 | 8.54 | **86.9** | | **Qwen-RobotNav-4B** | **90.0** | 6.40 | 77.4 | | **Qwen-RobotNav-8B** | 89.7 | 5.70 | 78.6 | Embodied Question Answering Equipped with the agentic system, Qwen-RobotNav sets new state-of-the-art results across three EQA benchmarks, surpassing prior methods by substantial margins. | Method | HM-EQA Acc.↑ | MT-EQA Acc.↑ | EXPRESS LLM Score↑ | | ----------------------------- | ------------ | ------------ | ------------------ | | Explore-EQA | 58.4 | 36.2 | — | | Memory-EQA | 61.4 | 43.1 | — | | FAST-EQA | 69.2 | 50.5 | 68.7 | | **Qwen3.5-Plus + QwenNav-8B** | 74.1 | 52.1 | 77.66 | | **Qwen3.6-Plus + QwenNav-8B** | **76.7** | **54.4** | **79.27** | Autonomous Driving (NAVSIM) On closed-loop driving evaluation, Qwen-RobotNav-4B achieves **91.4 PDMS**, surpassing specialized driving models. | Method | NC↑ | DAC↑ | TTC↑ | Comf.↑ | EP↑ | PDMS↑ | | -------------------- | -------- | -------- | -------- | ------- | -------- | -------- | | NavFoM | 97.7 | 93.5 | 92.3 | **100** | 79.6 | 84.3 | | AutoVLA | 98.4 | 95.6 | 98.0 | 99.9 | 81.9 | 89.1 | | ReCogDrive | 97.9 | 97.3 | 94.9 | **100** | **87.3** | 90.8 | | ReflectDrive | 97.7 | **99.3** | 93.5 | **100** | 86.9 | 91.1 | | **Qwen-RobotNav-4B** | **99.8** | 97.5 | **98.5** | 99.9 | 84.4 | **91.4** | | **Qwen-RobotNav-8B** | **99.8** | 96.9 | 98.2 | 99.9 | 84.2 | 90.9 | Qwen-RobotNav produces temporally consistent curved trajectories across both NAVSIM and the AlpaSim closed-loop simulator (zero-shot). * * * ## Real-World Deployment We deploy Qwen-RobotNav on a Unitree Go2 quadruped robot with on-device inference via NVIDIA Jetson Thor, achieving **196ms latency (5.1 Hz)**. The sole visual input is the Go2’s built-in low-resolution camera. All experiments are conducted zero-shot in previously unseen environments without any environment-specific fine-tuning. Deployment The robot executes navigation tasks in an apartment setting using step-by-step verbal instructions, traversing between the bedroom, living room, and bathroom while responding to fine-grained spatial directives. Instruction Following We evaluate a back-and-forth navigation task in an unseen exhibition hall: the robot first navigates 21.78 m from a living room to a hospital room following language instructions, then receives a reverse command and must precisely retrace the entire route. This is particularly challenging as it requires the model to maintain spatial awareness over long distances, ground diverse visual landmarks in both forward and reverse directions, and execute accurate bidirectional position control purely from language. Agentic Navigation In agentic mode, the system supports open-ended requests beyond simple route-following. Given the instruction "check whether a green umbrella was left at Cotti Coffee," the agent decomposes the task into sub-goals, navigates using corridor landmarks for localization, inspects the target scene, and produces an evidence-grounded answer without human intervention. #### Agentic navigation: autonomously checking for a green umbrella at Cotti Coffee. * * * ## What’s Next Qwen-RobotNav reframes multi-task navigation as a context modeling problem: diverse tasks share the same perception and planning backbone but require different strategies for consuming observations. Treating context as an externally controllable interface enables a single model to serve as a practical, deployable navigation primitive within agentic systems. [← Back to Qwen-Robot Suite](https://qwen.ai/blog?id=qwen-robotsuite) * * * ## Citation ```bibtex @article{qwenrobotnav2026, title={Qwen-RobotNav: A Scalable Navigation Model Designed for an Agentic Navigation System}, author={Qwen Team}, year={2026} } ```