ReflectVLN: Training Vision-Language Navigation Agents with Reflective Reasoning

Closed-loop Intention–Execution Agents for Long-Horizon VLN

Jiahang Wang^1*, Yirong Yang^2*, Yanqing Zhu^1†, Minghua Luo¹, Shichao Xie¹, Fei Liu¹, Mu Xu¹

¹Amap, Alibaba Group    ²Beihang University
^*Equal contribution    ^†Project lead

This repository is the official implementation of ReflectVLN, an agentic Vision-and-Language Navigation (VLN) framework with bidirectionally interactive intention and execution agents. The execution agent tracks sub-goal progress and emits deviation signals to trigger on-demand reflection; the intention agent returns structured language guidance for recovery. With Action Chain-of-Thought (Action-CoT) and a reflection-driven data pipeline, ReflectVLN achieves competitive success and path efficiency on R2R-CE / RxR-CE under a constrained data budget, using only monocular RGB observations.

✨ Highlights

1. Closed-loop intention–execution interaction. Unlike one-way slow–fast / dual-system cascades, the execution agent emits progress and deviation signals (<VLNBOR>, <VLNBOC>, <VLNBOA>) to trigger the intention agent on demand, enabling timely sub-goal updates and recovery.

2. Action Chain-of-Thought (Action-CoT). A path-conditioned dual-query scheme first predicts a coarse future route, then conditions short-horizon waypoint control—improving temporal coherence with interpretable intermediate decisions.

3. Reflection-driven data pipeline. One-round failure rollouts are converted into verified error-correction trajectories with reflective language feedback (~100k samples), without multi-round iterative flywheels or large-scale ScaleVLN pretraining.

4. Strong results under a tight budget. Two independently parameterized Qwen2.5-VL-3B agents (~1.6M samples from R2R+RxR+reflection) reach 62.8% SR / 58.5 SPL on R2R-CE and 66.0% SR / 57.2 SPL on RxR-CE (Val-Unseen), with fewer high-level intention calls (~once every 9.63 steps).

✅ TODO

• Release the project page and paper PDF
• Release reflection dataset
• Open-source training and inference code
• Release pretrained model checkpoints
• Add installation and environment setup instructions

🧭 Method Overview

ReflectVLN organizes navigation as bidirectional interaction between two agents (same Qwen2.5-VL-3B architecture, independent parameters):

Agent	Role	Output
Intention agent $\pi_{\theta_{\mathrm{int}}}$	High-level reflective planner	Language sub-goal $c_t$ (regular or corrective)
Execution agent $\pi_{\theta_{\mathrm{exe}}}$	Short-horizon controller + monitor	Waypoints $\hat{A}_t$, coarse route $\hat{P}_t$, status token $z_t$

Status tokens control the system mode:

• <VLNBOA> — continue action execution (no intention call)
• <VLNBOR> — sub-goal completed; request next regular sub-goal
• <VLNBOC> — off-track; request corrective reflection

Two-stage training:

1. Stage I — Action-CoT pretraining: train the execution agent on expert trajectories (route + short-horizon losses).
2. Stage II — Status-aware fine-tuning: jointly train intention (language imitation) and execution (waypoints + status tokens) on expert + reflective recovery data.

Observations are monocular RGB only (current + history frames)—no panorama, depth, or odometry at inference.

📊 Results

Comparison on VLN-CE (Val-Unseen)

Method	Obs.	R2R NE↓	R2R OSR↑	R2R SR↑	R2R SPL↑	RxR NE↓	RxR SR↑	RxR SPL↑
StreamVLN	S.RGB	4.98	64.2	56.9	51.9	6.22	52.9	46.0
CorrectNav*	S.RGB	4.50	—	61.4	59.0	4.40	63.1	57.0
DualVLN	S.RGB	4.05	70.7	64.3	58.5	4.58	61.4	51.8
ReflectVLN (Ours)	S.RGB	4.19	67.3	62.8	58.5	3.98	66.0	57.2

*CorrectNav* denotes the first-round self-correction setting (comparable data regime).

Training efficiency

Method	Backbone	Data sources	Samples
CorrectNav	7B	R2R+RxR+ScaleVLN+VQA	~3M (1st round)
DualVLN	7B	R2R+RxR+ScaleVLN	~12M
ReflectVLN	2×3B	R2R+RxR+reflection	~1.6M

🔒 License

This project is released under the Apache 2.0 license. See LICENSE for details. Matterport3D / Habitat assets are subject to their original licenses.

✏️ Citation

If you find ReflectVLN useful in your research, please consider citing:

@article{wang2026reflectvln,
  title   = {ReflectVLN: Training Vision-Language Navigation Agents with Reflective Reasoning},
  author  = {Wang, Jiahang and Yang, Yirong and Zhu, Yanqing and Luo, Minghua and Xie, Shichao and Liu, Fei and Xu, Mu},
  journal = {arXiv preprint arXiv:2607.12680},
  year    = {2026}
}

🤝 Contact

For questions and feedback, please open a GitHub Issue or contact the authors.