Towards Simple and Frontier
Vision-Language Navigation

Lingyi Hong^1,*,†, Weiren Cai^2,*, Jiayu Chen^1,*, Xinyu Zhou^2,†,‡, Bailin He¹, Yunfei Guo¹, Qiao Sun³, Xiang Liu¹, Boyu Ma², Jianfei Yang², Wenqiang Zhang¹

¹ Fudan University ² Nanyang Technological University ³ Lenovo Group Limited

* Equal contribution · † Project leaders · ‡ Corresponding author

How far can a native multimodal large language model go in VLN before introducing task-specific architecture?

arXiv GitHub Models

76.7%R2R-CE · val-unseen SR

80.1%RxR-CE · val-unseen SR

293msModel-only latency · A100 · Qwen3-VL 4B

**Motivation and performance comparison of SimpleVLN.** (a) Motivation for using native MLLMs in VLN. (b) Comparison with existing methods on R2R-CE and RxR-CE.

Exploring what matters for vision-language navigation.

Formulation Data Optimization

Abstract.

SimpleVLN explores how far native multimodal models can go in vision-language navigation through better formulation, data, and optimization, without task-specific architectural designs.

Read the full abstract

Vision-and-Language Navigation (VLN) requires embodied agents to ground language in visual observations and make long-horizon decisions. Existing methods often address these demands through task-specific architectures. As multimodal large language models (MLLMs) become increasingly capable, an important question remains: Have we truly unlocked the potential of native MLLMs for VLN before introducing task-specific architecture?

We present SimpleVLN, an architecture-preserving framework developed through studies of navigation formulation, data construction, and optimization. At the formulation level, SimpleVLN combines single-turn decision inputs with fixed-budget global–local memory to efficiently use navigation history without unbounded context growth, while action chunks enable efficient output of multiple primitive actions per model invocation. For data construction, we combine expert demonstrations with policy-dependent DAgger experience and develop a scalable synthesis pipeline that generates large-scale tasks for reinforcement learning. At the optimization level, a progressive IL–DAgger–RL pipeline advances from action imitation to task-level outcome optimization.

Comprehensive experiments validate the effectiveness of these three aspects. We instantiate SimpleVLN across multiple MLLM backbones and model sizes, demonstrating its generality. SimpleVLN achieves 76.7% and 80.1% success rates on the R2R-CE and RxR-CE splits, respectively, substantially outperforming existing methods. These results show that the navigation capabilities of native MLLMs can be effectively unlocked without task-specific architectural designs.

01 — Demos

Qualitative
results.

We evaluate SimpleVLN on the R2R-CE and RxR-CE val-unseen benchmarks in Matterport3D with Habitat. We further evaluate SimpleVLN in real environments using a Daystar quadruped robot. The model runs onboard on an NVIDIA Jetson Thor GPU. SimpleVLN demonstrates strong navigation performance in unseen environments and follows multi-step instructions in diverse real-world scenes.

Simulation

Foyer

Living room

Gallery

Kitchen

Real world

Office hallwayTurn left and enter the first room

Yellow pillarStop next to the yellow pillar

KFC entranceWalk to the KFC entrance and stop

Vending machineStop at the vending machine

Outdoor pathFollow the stone path to the fitness area

Red carAvoid the pillar and stop at the red car

Follow the personSpot the person in black and follow

02 — Approach

Adapting native
MLLMs to VLN.

SimpleVLN directly adapts off-the-shelf MLLMs to VLN without architectural modifications. We study native-MLLM navigation from three perspectives: navigation formulation, data construction, and a progressive optimization strategy that shifts from action imitation to outcome-driven learning through environment interaction.

**Overview of SimpleVLN.** At each decision, sampled history frames, current observation, and the instruction are fed to an MLLM, which directly predicts an action chunk for navigation.

01Formulation

Decision inputs
and history management.

We combine single-turn decision inputs with fixed-budget global–local memory to efficiently use navigation history without unbounded context growth. Action chunking predicts multiple primitive actions per model invocation.

Single-turn decisions
Global–local memory
Action-chunk prediction

02Data

Demonstrations and
synthesized tasks.

We combine expert demonstrations with policy-dependent experience collected through DAgger, and develop a scalable task-synthesis pipeline to generate navigation tasks for reinforcement learning.

R2R + RxR + ScaleVLN
Policy-dependent DAgger
Scalable task synthesis

03Optimization

Progressive
IL–DAgger–RL training.

We investigate a progressive IL–DAgger–RL pipeline: imitation learning initializes the policy, DAgger adapts it to policy-induced states, and RL optimizes task-level outcomes through environment interaction.

Imitation learning
DAgger-based fine-tuning
Outcome-driven RL

Training pipeline of SimpleVLN — **Training data pipeline of SimpleVLN.** SimpleVLN first learns from expert demonstrations, then collects and aggregates corrective trajectories through DAgger. For RL, it synthesizes navigation episodes from sampled start–goal paths and generated instructions, then optimizes the policy through environment interaction and outcome feedback.

03 — Results

Quantitative
results.

SimpleVLN achieves state-of-the-art navigation success while preserving the native MLLM architecture and using only first-person RGB observations. Strong results across model families and scales show that this approach can unlock navigation capabilities in a range of native multimodal models.

R2R-CEval-unseen · SR ↑

76.7%

Qwen3.5 9B

Qwen3-VL 4B73.8
Qwen3-VL 8B74.4
Qwen3.5 4B76.3

RxR-CEval-unseen · SR ↑

80.1%

Qwen3.5 9B

Qwen3-VL 4B75.1
Qwen3-VL 8B76.4
Qwen3.5 4B79.2

Inference latencyNVIDIA A100

293ms

Qwen3-VL 4B

SimpleVLN predicts each action chunk in 293 ms with Qwen3-VL 4B on an NVIDIA A100, while maintaining strong navigation without a slower model or task-specific architecture.

Comparison with state-of-the-art methods on the R2R-CE and RxR-CE val-unseen splits.

This comparison covers navigation accuracy, path efficiency, and observation requirements. SimpleVLN achieves strong performance using only first-person RGB, without task-specific architectural modifications. Consistent gains across backbones support the applicability of the training recipe beyond a single model. † and * denote results before and after RL, respectively.

Method	Size	Observation				R2R-CE				RxR-CE
Method	Size	Pano.	Odo.	Depth	S.RGB	NE↓	OS↑	SR↑	SPL↑	NE↓	SR↑	SPL↑	nDTW↑
NaVid	7B					5.47	49.1	37.4	35.9	8.41	23.8	21.2	—
Uni-NaVid	7B					5.58	53.3	47.0	42.7	6.24	48.7	40.9	—
NaVILA	7B					5.22	62.5	54.0	49.0	6.77	49.3	44.0	58.8
ActiveVLN	3B					5.31	58.3	50.1	43.7	5.84	50.7	41.2	58.1
StreamVLN	7B					4.98	64.2	56.9	51.9	6.22	52.9	46.0	61.9
NavFoM	7B					5.01	64.9	56.2	51.2	5.51	57.4	49.4	60.2
JanusVLN	7B					4.78	65.2	60.5	56.8	6.06	56.2	47.5	62.1
NaVIDA	3B					4.32	69.5	61.4	54.7	5.23	57.4	49.6	67.0
CorrectNav	7B					4.24	67.5	65.1	62.3	4.09	69.3	63.3	75.2
DualVLN	7B					4.05	70.7	64.3	58.5	4.58	61.4	51.8	70.0
FutureNav	4B					4.24	70.2	65.4	61.3	4.60	60.9	52.8	68.0
FutureNav	8B					4.29	69.2	64.3	60.3	4.26	63.9	54.8	69.1
AgentVLN	3B					3.88	73.5	67.2	64.7	3.92	69.5	61.3	74.6
OmniNav	3B					3.74	74.6	69.5	66.1	3.77	73.6	62.0	—
SEDualVLN	7B					3.75	73.7	67.3	62.5	4.12	63.9	52.4	72.8
Qwen-RobotNav	4B					4.22	73.6	66.9	60.5	4.15	71.3	61.5	68.6
Qwen-RobotNav	8B					4.36	72.7	65.7	59.6	4.16	73.4	63.5	69.9
Qwen-RobotNav	4B					3.80	77.2	69.5	63.6	3.80	75.2	65.0	71.9
Qwen-RobotNav	8B					3.53	78.5	72.1	66.6	3.58	76.5	65.7	72.5
Efficient-VLN	8B					3.29	78.2	73.2	65.0	3.18	75.6	60.0	—
SimpleVLN† (Qwen2.5-VL)	3B					5.70	63.3	51.6	44.9	5.20	50.3	42.0	57.6
SimpleVLN† (Qwen2.5-VL)	7B					4.87	68.4	59.1	52.0	5.31	58.9	49.2	62.9
SimpleVLN† (Qwen3-VL)	4B					4.26	74.3	65.4	58.8	4.79	63.4	54.1	66.5
SimpleVLN* (Qwen3-VL)	4B					3.39	79.4	73.8	66.7	3.29	75.1	65.3	71.5
SimpleVLN† (Qwen3-VL)	8B					4.40	73.3	64.2	57.0	4.73	64.2	54.3	66.2
SimpleVLN* (Qwen3-VL)	8B					2.93	79.2	74.4	69.5	3.11	76.4	66.4	77.4
SimpleVLN† (Qwen3.5)	4B					3.76	77.5	67.1	59.2	4.32	65.6	55.3	67.8
SimpleVLN* (Qwen3.5)	4B					2.79	80.1	76.3	72.3	2.61	79.2	69.7	80.3
SimpleVLN† (Qwen3.5)	9B					3.59	76.9	69.2	61.2	4.48	64.0	54.2	67.0
SimpleVLN* (Qwen3.5)	9B					2.63	84.1	76.7	72.6	2.51	80.1	69.9	79.8

Comparison with state-of-the-art methods on inference latency and navigation performance.

Latency is measured on R2R-CE using the same NVIDIA A100 GPU, and measures model execution only for one action-chunk decision, excluding preprocessing, environment execution, and other non-model overhead. SimpleVLN uses Qwen3-VL-4B as its backbone.

SimpleVLN combines high navigation success with low model latency. The comparison shows that stronger navigation does not require a slower decision model or additional task-specific architecture.

Method	Size	Latency (ms)↓	R2R-CE				RxR-CE
Method	Size	Latency (ms)↓	NE↓	OS↑	SR↑	SPL↑	NE↓	SR↑	SPL↑	nDTW↑
StreamVLN	7B	300	4.98	64.2	56.9	51.9	6.22	52.9	46.0	61.9
JanusVLN	7B	906	4.78	65.2	60.5	56.8	6.06	56.2	47.5	62.1
DualVLN	7B	683	4.05	70.7	64.3	58.5	4.58	61.4	51.8	70.0
FutureNav	4B	1141	4.24	70.2	65.4	61.3	4.60	60.9	52.8	68.0
SimpleVLN	4B	293	3.39	79.4	73.8	66.7	3.29	75.1	65.3	71.5

04 — Analysis

Navigation
analysis.

We investigate how navigation formulation, data construction, and optimization affect VLN performance. The following experiments use controlled ablations of these three aspects.

Formulation

Navigation interface and memory design.

(a) Action chunking. Action chunking improves navigation performance, supporting it as an effective action-output representation rather than merely a way to reduce model invocation frequency.

(b) Single-turn navigation. Reconstructing each decision from the current observation and sampled visual history performs better than continually accumulating interaction history in multi-turn dialogue. Although multi-turn dialogue can preserve a fuller interaction history, continually appending observations and actions creates an overly long context that current MLLMs cannot effectively exploit for navigation.

(c) Observation sampling. Uniformly sampling global-history observations performs better than sampling only at action-chunk boundaries or retaining all primitive-step observations, favoring compact, uniformly sampled global history over retaining every frame.

Action format
Action format	NE↓	OS↑	SR↑
Primitive action	5.2	54.5	45.6
Action chunk	5.5	61.2	53.6

Dialogue format
Dialogue format	NE↓	OS↑	SR↑
Multi-turn	5.9	55.2	35.7
Single-turn	5.5	61.2	53.6

Observation sampling
Sampling scheme	NE↓	OS↑	SR↑
Chunk	6.1	55.2	46.3
Primitive (all)	5.8	57.5	51.9
Primitive (8)	5.5	61.2	53.6

Global memory budget — **Memory budget.** Effect of global and local memory sizes on R2R-CE val-unseen SR. We vary M with N = 1 (left) and N with M = 8 (right); stars indicate the default configuration.

Local memory budget — **Memory budget.** Effect of global and local memory sizes on R2R-CE val-unseen SR. We vary M with N = 1 (left) and N with M = 8 (right); stars indicate the default configuration.

Data

Demonstration and DAgger data on the R2R-CE and RxR-CE val-unseen splits.

This table separates the contribution of broader expert demonstrations from experience collected by the learned policy. Combining demonstration sources improves navigation, but DAgger brings the largest additional gain. Expert routes alone do not cover all states the robot encounters during its own execution; policy-dependent trajectories provide valuable complementary supervision.

R2R SR↑	R2R SPL↑	RxR SR↑	RxR SPL↑
53.6	48.1	50.3	42.0
56.7	51.3	54.5	46.7
57.9	50.9	55.3	48.1
65.4	58.8	63.4	54.1

Synthesized RL data on the R2R-CE and RxR-CE val-unseen splits.

Synthesizing RL tasks from a mixture of R2R, RxR, and ScaleVLN scene pools performs best, suggesting that combining scene sources provides broader training coverage than relying on any single pool.

R2R	R2R SR↑	R2R SPL↑	RxR SR↑	RxR SPL↑
Pre-RL baseline	65.4	58.8	63.4	54.1
	68.5	63.2	65.8	55.3
	67.5	59.3	68.5	59.5
	66.6	61.3	66.3	56.2
	69.4	63.8	69.7	60.8
	70.1	65.6	71.0	62.1

Optimization

RL reward curve — **RL training dynamics.** Rollout reward, validation success, and the average number of action-chunk decisions track how the policy changes during environment interaction.

RL success-rate curve — **RL training dynamics.** Rollout reward, validation success, and the average number of action-chunk decisions track how the policy changes during environment interaction.

Reward design and RL optimization on the R2R-CE and RxR-CE val-unseen splits with Qwen3-VL 4B.

Binary task-completion feedback achieves higher success rates on both benchmarks than an SPL-based reward, supporting simple task-completion feedback as an effective RL signal.

The evaluated DAPO configuration further improves navigation with binary success rewards. However, since the compared configurations differ in more than the optimizer, the gain cannot be attributed solely to the optimization algorithm.

RL configuration	R2R SR↑	R2R SPL↑	RxR SR↑	RxR SPL↑
Pre-RL baseline	65.4	58.8	63.4	54.1
GRPO + SPL(τ)	68.9	63.1	67.7	58.3
GRPO + 1(success)	70.1	65.6	71.0	62.1
DAPO + 1(success)	71.5	65.3	73.7	64.0

BibTeX

@article{hong2027simplevln,
  title   = {SimpleVLN: Towards Simple and Frontier
             Vision-Language Navigation},
  author  = {Hong, Lingyi and Cai, Weiren and Chen, Jiayu
             and Zhou, Xinyu and He, Bailin and Guo, Yunfei
             and Sun, Qiao and Liu, Xiang and Ma, Boyu
             and Yang, Jianfei and Zhang, Wenqiang},
  year    = {2027},
  eprint  = {2610.23203},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url     = {https://arxiv.org/abs/2610.23203}
}

Towards Simple and FrontierVision-Language Navigation