VLA Gallery / All models / Navigation

Qwen-RobotNav

Qwen Team · 2026-06

VLANavigation4B · 8Bregression

Parameterizes task mode and observation context (token budget·temporal decay·camera weighting) and randomizes them during training, so one model performs VLN·object search·tracking·autonomous driving as tools for a high-level agent planner.

Paper

Qwen-RobotNav Technical Report: A Scalable Navigation Model Designed for an Agentic Navigation SystemarXiv 2606.18112

Overview

Category

Navigation

Type

VLA

Released

2026-06

Organization

Qwen Team

Open weights

No

Parameters

4B · 8B (based on Qwen3-VL; no sentence giving the total, abstract mentions 2B–8B scaling)

Architecture

Backbone LLM

Qwen3-VL-8B (including 4B variant)

Vision encoder

SigLIP-2 ViT (built into Qwen3-VL, DeepStack)

Action head

4-layer MLP waypoint regression (hidden 512) (regression)

Chunk

8 waypoints

Control rate

4.9–5.1 Hz (Go2 real deployment)

Architecture (original summary)

Qwen3-VL (4B·8B; SigLIP-2 ViT·2D-RoPE dynamic resolution·DeepStack) + 4-layer MLP action head (hidden 512) → continuous regression of 8 waypoints (x,y,θ) (MSE); task mode (VLN·PointNav·ObjNav·Tracking), visual token budget (2048–4096), temporal decay, camera weighting and frame sampling adjusted at inference; called like a tool by a higher-level planner; about 5 Hz in real deployment

Training

Pretraining accelerators

H100 (count not disclosed)

Pretraining duration

Not disclosed

Pretraining steps & batch

8B global batch 256, lr 2e-5 (ViT · LLM) · 1e-4 (action head); steps and epochs not disclosed

GPU/TPU-hours

2,816 (Stated · H100-hours ("a total of 2,816 H100 GPU hours", 8B model))

Fine-tuning

N/A (single-stage end-to-end training from Qwen3-VL)

Data type

Simulation·synthetic / real-robot autonomous driving logs / web vision-language

Data scale

15.6M samples (trajectories 85% · VL 15%): instruction following 5.63M (R2R 1,491K · RxR 4,140K), PointNav 984K, ObjectNav 2,000K, tracking 1,486K (EVT-Bench), autonomous driving about 3.2M (nuScenes 78K · OpenScene 3,138K), video generator data 40K; general VL about 1.0M + navigation reasoning about 873K

Main datasets

VLN-CE R2R·RxR; Matterport3D·HM3D (Habitat); EVT-Bench; nuScenes; OpenScene (NAVSIM); NavGPT-2-style reasoning data; general VL such as VQAv2·RefCOCO

Tasks & evaluation

R2R-CE val-unseen SR 72.1%·SPL 66.6 (8B panoramic; monocular 8B 65.7%·4B 66.9%), RxR-CE SR 76.5%·SPL 65.7; HM3Dv2 ObjectNav SR 75.6% (4B); EVT-Bench TR 90.0% (4B); NAVSIM PDMS 91.4 (4B); agentic EQA HM-EQA 76.7% (Qwen3.6-Plus planner); real-robot Go2 zero-shot 21.78 m round-trip VLN

Robot

Unitree Go2 quadruped (uses only built-in low-resolution camera)

Inference

Inference device & speed

Unitree Go2 (4B model): remote server average 196 ms (5.1 Hz), Jetson Thor on-device (FP8·TensorRT) 204 ms (4.9 Hz) (paper); README lists Jetson Thor on-device 196 ms (5.1 Hz); memory not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

16 GB

Where it runs

Onboard Jetson Thor (FP8·TensorRT) or remote server

Reinforcement learning & base

Reinforcement learning

Not used · Trajectory MSE + VL next-token supervised learning (trajectory:VL 85:15 co-training)

Base control

Quadruped (Waypoints → tracking controller) · Model outputs 8 waypoints (x,y,θ) → tracked by Go2 (low-level controller Not stated); a high-level planner calls it with different task mode·context settings

Notes

Key features

Parameterizes task mode and observation context (token budget·temporal decay·camera weighting) and randomizes them during training, so one model performs VLN·object search·tracking·autonomous driving as tools for a high-level agent planner. The 8B trained on 15.6M samples reached R2R-CE SR 72.1%·RxR-CE SR 76.5% (panorama) and drove zero-shot on Go2 at 4.9 Hz on-device on Jetson Thor.

Notes

README: 'There is currently no plan to release the model weights'. Conflicting latency figures: paper gives remote 196 ms·on-device 204 ms, README gives on-device 196 ms. Monocular result (R2R SR 65.7%) matches the LightNav-0 comparison table. 'Real-robot autonomous driving logs' are vehicle driving data (nuScenes·OpenScene). The 2B variant in the abstract is absent from the results tables.

Verification

Confirmed