QwenLM/Qwen-RobotNav - 仓库详情

AI 仓库情报摘要

FR-AI / ANALYSIS

为什么值得关注

该模型在 VLN-CE、物体搜索、跟踪和自动驾驶等多个导航基准上取得了领先结果，并展示了在 Unitree Go2 机器人上零样本部署的能力（在 Jetson Thor 上以 5.1 Hz 运行）。其可调节的观测协议使其成为上层智能体系统的理想导航基元。

适合谁使用

从事导航和具身智能研究的机器人学者
构建多任务机器人控制器的自主系统工程师
关注可扩展导航模型的计算机视觉与 VLM 研究者
需要可重配置航点预测器的智能体系统开发者

典型使用场景

通过自然语言指令进行室内导航
在未知环境中进行物体搜索与目标跟踪
基于多视角输入的自动驾驶仿真
需要探索与证据收集的具身问答任务

项目优势

单一模型统一覆盖 VLN、点导航、物体导航、跟踪、自动驾驶和具身问答等多种任务，无需针对每个任务微调
可控制的观测协议（视觉令牌预算、时间衰减、相机权重）允许在推理时调整模型行为以适应不同任务
在 1560 万个样本上进行训练，结合轨迹监督与视觉-语言协同训练，从 2B 到 8B 参数表现出良好的扩展性
零样本部署至 Unitree Go2 机器人，在 Jetson Thor 上实现 5.1 Hz 的推理速度

使用前须知

模型权重尚未公开发布，目前没有公布计划
真实世界验证仅限于单一机器人平台（Unitree Go2）和特定场景
观测配置中的令牌预算与时间衰减需要在长程记忆与近时细节之间进行权衡，消融实验显示了这些权衡

README 快速开始

Qwen-RobotNav

A Scalable Navigation Model Designed for an Agentic Navigation System

Qwen Team

📑 Technical Report |
📖 Blog |
🖥️ Demo

Welcome to the official repository of Qwen-RobotNav. Here, you can find official information about Qwen-RobotNav and post your questions (Issues).

Note: There is currently no plan to release the model weights for Qwen-RobotManip or Qwen-RobotNav. We will continue adding report resources that can be publicly released to this repository.

🎬 Demo

If the video does not render in your browser, open the direct demo preview link. For the full-length high-resolution source, open big_agent.mp4.

Feature Highlights

This blog demo highlights the key design features of Qwen-RobotNav: unified multi-domain navigation, controllable observation context, agentic tool-call style execution, and zero-shot real-world deployment.

If the feature video does not render in your browser, open the direct feature preview link. For the full-length high-resolution source, open Nav_blog_demo.mov.

💡 Introduction

Qwen-RobotNav is a scalable navigation model built on Qwen3-VL. It unifies instruction following, point-goal and object-goal navigation, target tracking, autonomous driving, and embodied question answering under a shared waypoint-prediction interface.

The key idea is to treat navigation as context modeling. Different navigation tasks share a perception-planning backbone, but they require different strategies for consuming visual history: long-horizon instruction following needs memory, target tracking needs recent high-resolution frames, object search shifts between exploration and local approach, and driving depends on multi-view short-term motion context.

Qwen-RobotNav exposes this difference as a configurable observation protocol. An upper-level planner can call the same model with different task modes and context parameters, making Qwen-RobotNav a natural navigation primitive for

项目描述

Official Repo for Qwen-RobotNav