VLA Gallery / All models / Navigation

VLN-R1

The University of Hong Kong·Shanghai AI Laboratory · 2025-06

VLANavigation2B / 7Bdiscrete tokens

Applies SFT to a general VLM (Qwen2-VL) for egocentric-video-based VLN, then improves multi-step action prediction via GRPO reinforcement fine-tuning (RFT) with a time-decayed reward (TDR).

Paper

VLN-R1: Vision-Language Navigation via Reinforcement Fine-TuningarXiv 2506.17221

Overview

Category

Navigation

Type

VLA

Released

2025-06

Organization

The University of Hong Kong·Shanghai AI Laboratory

Open weights

No

Parameters

2B / 7B (Qwen2-VL-2B·7B)

Architecture

Backbone LLM

Qwen2-VL-7B / 2B (built-in LLM name not stated)

Vision encoder

Qwen2-VL built-in vision encoder (details not stated)

Action head

Text of the next 6 discrete actions (option letter + description) (discrete tokens)

Chunk

6 actions

Control rate

Not disclosed

Architecture (original summary)

Full fine-tuning of Qwen2-VL-2B/7B; first-person video (16 images per instance, up to 65,536 pixels, about 4.1K tokens) + instruction input; outputs the next 6 discrete actions as option letter+description text (FORWARD·TURN-LEFT·TURN-RIGHT·STOP, e.g. 'C. Turn right 30 degrees')

Training

Pretraining accelerators

8 × A800 (7B SFT, DeepSpeed ZeRO-3)

Pretraining duration

36 hours (SFT 1 epoch)

Pretraining steps & batch

SFT 1 epoch · global batch 64 (2 per GPU × 4 accumulation, lr 5e-6)

GPU/TPU-hours

288 (Converted · 8× A800 × 36 hours (7B SFT 1 epoch))

Fine-tuning

RFT (GRPO): about 12 hours per epoch (epoch count not stated), lr 1e-6, 8 samples per prompt, KL β 0.04, batch 1 per GPU, 20K samples (7B on 8 A800 setup)

Data type

Simulation·synthetic

Data scale

VLN-Ego: R2R 630K + RxR 1.2M training samples (SFT 1.8M); RFT 10K per dataset, 20K total

Main datasets

VLN-Ego (generated from Habitat·MP3D; based on R2R 7,189 paths·RxR 42,023 trajectories)

Tasks & evaluation

R2R-CE val-unseen SR 30.2%·SPL 21.8% (7B), SR 25.6%·SPL 20.5% (2B); RxR-CE SR 22.3% (7B)·21.4% (2B); cross-domain (R2R SFT→RxR RFT) RxR SR 22.7% (7B)

Robot

No real robot (Habitat sim agent)

Inference

Inference device & speed

Not disclosed (Habitat sim evaluation only)

Inference memory

Not disclosed

Weights bf16 (calc.)

14 GB

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Used (VLA RL post-training) · GRPO-based RFT stage after SFT (multi-step action reward with Time-Decayed Reward)

RL evidence

https://arxiv.org/abs/2506.17221

Base control

Sim only (Discrete motion actions) · No real robot: a Habitat VLN-CE sim agent executes discrete actions (forward·turn·stop)

Notes

Key features

Applies SFT to a general VLM (Qwen2-VL) for egocentric-video-based VLN, then improves multi-step action prediction via GRPO reinforcement fine-tuning (RFT) with a time-decayed reward (TDR). The 2B model after RFT outperformed a 7B model with SFT only, showing the effect of RL post-training even for small VLMs.

Notes

Weights not released (only VLN-Ego training data·data generation code released; SFT·RFT·evaluation code in preparation); 2B training hardware not stated; some figures inconsistent across tables (Table 1 vs Table 3); no real-robot experiments (stated as a limitation).

Verification

Confirmed