- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
SimpleVLN adapts a multimodal language model to vision-language navigation without modifying its architecture. It combines single-turn decision inputs, a fixed budget of global and local visual history, and action-chunk prediction with imitation learning, DAgger, and reinforcement learning. The Qwen3.5-9B model achieves 76.7% and 80.1% success on the R2R-CE and RxR-CE val-unseen splits.
Critical strengths
The paper presents an effective navigation training recipe with a simple architecture. Bounded visual history prevents input size from growing indefinitely with trajectory length, while action chunks reduce the frequency of model calls. Neither requires an additional task-specific network module.
The experiments systematically examine input format, history sampling, demonstration data, DAgger, and reinforcement learning. DAgger adds 7.5 and 8.1 percentage points in Table 4, and RL improves several model configurations in Table 1. These results provide useful empirical guidance beyond reporting the best-performing configuration.
The final results are competitive under single-RGB observations. Table 2 also compares model execution latency on a common A100, reporting 293 ms per action-chunk decision for Qwen3-VL-4B.
Critical weaknesses
The DAgger implementation is the most important missing detail. Section 3.3 describes retaining a policy trajectory or selecting an expert trajectory according to deviation from the reference, but does not specify the deviation measure, the switching point, or how supervision is obtained for retained policy trajectories. In particular, whether off-path states receive newly queried expert actions or the policy's own actions changes the interpretation of this training stage. Given the substantial gain attributed to DAgger, these details should be explicit. Some key details of synthetic RL data construction are missing. Section 3.3 describes sampling paths in training scenes and generating instructions from path waypoints, rooms, salient objects, and egocentric views. However, it does not identify the MLLM, provide the prompts or task counts, or explain whether quality filtering is applied. The scene-source ablations establish the value of different pools but do not establish the scale or quality of the synthetic data. Methods in Table 1 differ in backbones and training data, so the overall performance gap cannot be attributed directly to the training recipe. Section 4.1 already reports training epochs, learning rates, and an RL schedule of 40 steps with batch size 256 and 16 rollouts per prompt. The number of additional DAgger samples, total environment interactions, and training compute are still needed to assess the resources required for these gains. Finally, comparisons of history budgets and RL configurations are reported on val-unseen, but it is unclear whether this split also guides hyperparameter or checkpoint selection. Variability across training seeds is not reported either, making the smaller ablation differences difficult to assess.
Questions for authors
- What deviation measure and threshold does DAgger use? When a policy trajectory is retained, how are supervised action labels generated, particularly at states outside the reference path?
- Which MLLM and prompts generate the synthetic instructions, and is quality filtering applied? Please report the number of synthetic tasks and additional DAgger samples, total environment interactions, and training GPU-hours.
- Does each model use its last training checkpoint or one selected by validation performance? Is val-unseen used for checkpoint or hyperparameter selection? Have the key ablations been repeated across training seeds?
8 comments
Response to Reviewer 1xof - 2 / 5
W2: Section 3.3 describes sampling paths in training scenes and generating instructions from path waypoints, rooms, salient objects, and egocentric views. However, it does not identify the MLLM, provide the prompts or task counts, or explain whether quality filtering is applied. The scene-source ablations establish the value of different pools but do not establish the scale or quality of the synthetic data.
We thank the reviewer for pointing out that the description of our synthetic data construction lacks important implementation details.
MLLM and prompt. We use a MLLM (e.g. Qwen3.5 or GPT) with a fixed prompt to generate navigation instructions, which we will include verbatim in the appendix. At each node along a reference trajectory, the MLLM receives the ground-truth action, room label, instance-level landmark candidates (e.g., sofa#1), and the corresponding egocentric RGB frame. The prompt instructs the model to select a few salient landmarks rather than enumerate all candidates, merge consecutive actions, avoid hallucinating rooms, objects, or turns, and specify an explicit stopping location. We also provide up to eight human-written instructions from the same scene as style references. Since the generated instructions are grounded in the nodes and actions of the reference trajectory, they are constrained by the corresponding environment and navigation path.
Scale and quality control. We synthesize a total of 38,520 navigation tasks, including 5,450 from R2R, 7,070 from RxR, and 26,000 from ScaleVLN training scenes. The tasks are stratified into 50-step trajectory-length buckets. These are exactly the three data pools compared in Table 5.
We apply several filtering criteria to remove low-quality or unsuitable tasks. Specifically, we discard trajectories with fewer than 10 or more than 400 steps, geodesic lengths below 3 m, or fewer than two rooms. We also exclude tasks whose goals correspond to wall-mounted or unreachable objects, as well as trajectories that overlap with existing or real trajectories. In addition, we manually audit 50 randomly sampled tasks together with their reference trajectories, evaluating entity grounding, goal consistency, instruction clarity, and executability. Over 95% of the sampled tasks are judged to be reasonable and correct.
Pass@8-based filtering. Beyond these rule-based and manual quality checks, we further assess the difficulty and learnability of the synthesized tasks using the SFT-trained policy. Specifically, we evaluate each task with eight independent rollouts and retain tasks with Pass@8 scores between 0.2 and 0.8. This filtering removes tasks that are either too difficult for the current policy to make meaningful progress or too easy to provide sufficient learning signals. After this step, we retain 12,837 tasks in total, comprising 1,343 from R2R, 1,558 from RxR, and 9,936 from ScaleVLN. We will add these details and the final dataset statistics in Appendix.
Response to Reviewer 1xof - 3 / 5
W3: Methods in Table 1 differ in backbones and training data, so the overall performance gap cannot be attributed directly to the training recipe.
We thank the reviewer for raising this important concern. We agree that a direct comparison across methods with different backbones and training data cannot, by itself, establish the contribution of the training recipe. To address this issue, we have expanded our analysis in three aspects: (1) we explicitly report the backbone and training data of each method in Table 1; (2) we introduce matched-backbone comparisons to examine the gains from our progressive training pipeline; and (3) we provide a near-matched comparison with JanusVLN, as well as additional data-source ablations, to better understand the effects of architecture and training data.
1. Matched-backbone comparisons.
We first compare models using the same backbone and model size, which reduces the confounding effect of architectural differences. Table R1 presents the results for Qwen3-VL-4B. Our SFT model achieves performance comparable to existing methods with additional architectural components, while our RL-enhanced model obtains substantial further improvements without introducing any task-specific architectural modules.
Table R1. Matched-backbone comparison on Qwen3-VL-4B.
| Method | R2R-CE SR | RxR-CE SR | Extra modules | Training data |
|---|---|---|---|---|
| FutureNav-4B | 65.4 | 60.9 | VGGT + world-action heads | 2.43M samples |
| Qwen-RobotNav-4B (S.RGB) | 66.9 | 71.3 | MLP action head | 15.6M samples |
| SimpleVLN (SFT) | 65.4 | 63.4 | None | 2.71M samples |
| SimpleVLN (+RL) | 73.8 | 75.1 | None | +12,837 synthesized tasks |
Compared with FutureNav-4B, our SFT model achieves the same R2R-CE SR (65.4) and higher RxR-CE SR (63.4 vs. 60.9), despite using no additional spatial encoder or auxiliary action head. More importantly, incorporating our RL stage improves SR by 8.4 points on R2R-CE and 11.7 points on RxR-CE with the same backbone and model size. The RL stage uses 12,837 synthesized tasks without action annotations, allowing the policy to optimize task-level navigation success through environment interaction rather than additional expert action supervision.
The same trend is observed across other model configurations:
| Backbone and size | R2R-CE SR gain | RxR-CE SR gain |
|---|---|---|
| Qwen3-VL-4B | +8.4 | +11.7 |
| Qwen3-VL-8B | +10.2 | +12.2 |
| Qwen3.5-4B | +9.2 | +13.6 |
| Qwen3.5-9B | +7.5 | +16.1 |
These results suggest that the benefit of our RL-enhanced training pipeline is consistent across different backbone families and model sizes, rather than being specific to a particular architecture.
2. Near-matched comparison with JanusVLN.
We further compare SimpleVLN with JanusVLN, which uses the same Qwen2.5-VL-7B backbone and a closely related training corpus, including ScaleVLN and DAgger data. This provides a useful comparison for examining the effect of architectural modifications.
| Method | Backbone | ScaleVLN | DAgger | Extra module | R2R-CE SR | RxR-CE SR |
|---|---|---|---|---|---|---|
| JanusVLN | Qwen2.5-VL-7B | ~155K trajectories | ~14K trajectories | VGGT encoder | 60.5 | 56.2 |
| SimpleVLN (SFT) | Qwen2.5-VL-7B | ~155K trajectories | ~17K trajectories | None | 59.1 | 58.9 |
The two methods achieve comparable performance at the SFT stage, with SimpleVLN obtaining 59.1/58.9 SR compared with JanusVLN's 60.5/56.2. This comparison suggests that additional spatial architecture is not necessary to obtain competitive performance when using a comparable backbone and training corpus. It is worth noting that JanusVLN additionally uses a large number of image-text pairs for training, whereas SimpleVLN does not. Furthermore, our RL-enhanced models achieve substantial improvements without introducing such architectural components.
3. The RL gains are not simply explained by the ScaleVLN scene pool.
Since ScaleVLN contributes the largest portion of our training trajectories, we further examine whether the RL improvement is mainly caused by its scene coverage. As shown in Table 5, synthesizing RL tasks from ScaleVLN scenes alone yields a +1.2 SR improvement on R2R-CE, whereas using R2R-only scenes, RxR-only scenes, and their mixture yields improvements of +3.1, +2.1, and +4.7, respectively. The mixed-source setting performs best, suggesting that the benefit is associated with the diversity of synthesized tasks and the resulting outcome-based learning signals, rather than simply the volume of ScaleVLN scenes.
Overall, the results show that, under the same backbone, SimpleVLN achieves comparable performance using a similar or even smaller amount of training data. This suggests that the performance gains of SimpleVLN do not simply stem from more training data or a stronger backbone, but rather from improvements in the training recipe.
Response to Reviewer 1xof - 1 / 5
Thanks for your insightful feedback and support. We sincerely appreciate your valuable comments and your recognition of the novelty of our benchmark and the soundness of the experiments.
We follow the DAgger pipeline in StreamVLN. At each visited state, we query a shortest-path follower for an expert action toward the next waypoint on the reference path. Deviation is measured by cumulative disagreement between the candidate atomic action and this expert action. Let E denote the disagreement count and m=max(1,⌊ L_{ref}/(K-1)⌋), where L_{ref} is the reference action-sequence length and K is the number of reference waypoints. Expert intervention is triggered when E/m>0.8 or E>12. The switch occurs before executing the triggering action: we discard the remaining policy action sequence, reset the counter, and execute expert actions queried from the agent’s current state. Expert control continues until the next reference waypoint is reached, after which policy control resumes. Thus, correction occurs within the ongoing rollout.
Supervision uses the actions actually executed at the corresponding visited states. Steps controlled by the policy retain their executed policy actions as targets, including disagreements below the intervention threshold. During intervention, targets are the newly queried expert actions executed from those states. We do not retrospectively replace all retained policy actions with expert labels. This stage therefore combines supervision from retained policy behavior and online expert corrections.
Trajectory retention is a separate decision. We require a final goal distance below 0.5 m and a path-efficiency ratio r=dₛₕₒᵣₜₑₛₜ/max(dₛₕₒᵣₜₑₛₜ,ℓ_{executed}) below 0.93 for rollouts with a deviation-triggered intervention, or below 0.85 otherwise. Each episode is attempted up to four times, retaining the first rollout satisfying these criteria. Model-generated STOP commands are overridden before completion of the reference path.