Qwen-RobotNav: A Scalable Navigation Model Designed for Agentic Navigation Systems

Existing unified navigation models had the limitation of assuming only a specific memory scheme. To address this, Qwen-RobotNav treats context as an externally controllable variable, and is a scalable model built on top of Qwen3-VL.

This model optimizes navigation behavior at inference time through the following 4-Axis Observation Protocol.

• Visual token budget: total number of tokens across all cameras and timesteps
• Temporal decay: weighting of recent frames versus past frames
• Camera weights: importance settings per camera
• Frame sample mode: full history coverage or recent-frame-focused mode

Qwen-RobotNav achieves SOTA (State-of-the-art) across 5 domains—VLN, ObjNav, Tracking, Driving, and EQA—with a single set of weights. Notably, it shows consistent scaling performance from 2B to 8B parameters.

It is also designed as a two-stage system combined with the higher-level planner Qwen3.7-Plus, functioning as a core component of Agentic Navigation that decomposes complex goals into sub-goals and executes them. It has also been deployed zero-shot on the Unitree Go2 robot in real-world environments, demonstrating excellent generalization performance.