acceptodds
Under review as a conference paper at ICLR 2027

ConDenseDrive: Knowledge-Dense, Token-Sparse Visual Encoder for End-to-End Autonomous Driving

Abstract

End-to-end driving requires rich spatiotemporal representations to support scene understanding and trajectory planning under strict latency constraints. However, most systems initialize their visual encoders from generic image pretraining and densely process all patch tokens for each camera view and input frame, limiting both representation quality and inference efficiency. To address these challenges, we present ConDenseDrive, a knowledge-dense and token-sparse visual encoder for end-to-end autonomous driving. Our three-stage framework first makes the encoder knowledge-dense by distilling complementary visual, semantic, and geometric knowledge from three vision foundation models. It then adapts the distilled encoder to multi-frame end-to-end driving and integrates spatiotemporal context into register tokens for planning. Finally, it makes the encoder token-sparse by applying register-guided token compression, reducing redundant patch computation and yielding a 2.0× speedup while improving planning performance. ConDenseDrive achieves 94.5 PDMS on NAVSIM v1 and 90.8 EPDMS on NAVSIM v2, and reaches 37.4 HD-Score in zero-shot evaluation on HUGSIM. Extensive experiments and qualitative analyses demonstrate that visual encoder design provides an effective route to reliable and efficient end-to-end driving.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, Visual Representation Learning, Knowledge Distillation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_JBX4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors designed a new knowledge-dense and token-sparse visual encoder to extract visual, semantic, and geometric knowledge for better end-to-end autonomous driving. To solve the problem of spatiotemporal information redundancy, the authors design a register-guided token compression mechanism to reduce redundant patch computation across cameras and frames. The experiment shows that the visual encoder and the three-stage framework are effective.

Critical strengths

I agree that optimizing visual encoders is an effective method to enhance end-to-end planning capabilities, so I support the core claim of this paper: “makes the visual encoder knowledge-dense and token-sparse for end-to-end driving”. Considering the visual, semantic, and geometric knowledge required for autonomous driving, it is intuitive for the stage 1 of the framework to distill different information from multiple teachers. In addition, reducing patch computation is a practical way to improve inference efficiency. Finally, the author's experiments show that the framework achieves excellent results in different NAVSIM sets and HUGSIM.

Critical weaknesses

I think the paper would benefit from a direct quantitative evaluation of the distilled visual encoder’s semantic and geometric capabilities. The authors claim that the visual encoder in stage 1 transfers the semantic and geometric knowledge from the teachers. I believe that if the corresponding semantic and depth estimation performance can be evaluated, the overall paper will be more complete and the powerful capabilities of the visual encoder can be reflected.

Questions for authors
  1. What are the results of DriveZero on NAVSIM v1 set? The paper discusses DriveZero in the main text, but its results are not included in Table 1.
3 comments
AuthorsOpenReview ↗

Response to Reviewer JBX4

We appreciate the recognition of our contributions and address the raised points in detail below.

Question1:\color{red}{\mathit{Question}\\ 1:} I think the paper would benefit from a direct quantitative evaluation of the distilled visual encoder’s semantic and geometric capabilities.

Response1:\color{blue}{\mathit{Response}\\ 1:} Thank you for suggesting this quantitative evaluation. We evaluate semantic segmentation on Cityscapes and monocular depth estimation on KITTI. For each task, we train a separate prediction head of the same design on the frozen final-layer patch features of each encoder. For KITTI, we use 800 training pairs and 200 evaluation pairs from the selected validation set. We have added training and evaluation details to Appendix A.

EncoderCityscapes mIoU ↑KITTI RMSE ↓
DeiT-S27.017.800
DINOv3-S63.904.187
Distilled DeiT-S55.344.072

Relative to the original DeiT-S, the distilled encoder improves Cityscapes mIoU by 28.33 points and reduces KITTI RMSE by 47.8%. The distilled visual encoder captures richer semantic and geometric information. Under the same frozen-feature probing protocol, the distilled encoder scores lower than the comparably sized DINOv3 ViT-S/16 on Cityscapes but achieves lower KITTI RMSE. These results demonstrate the effectiveness of our multi-teacher visual distillation.

Question2:\color{red}{\mathit{Question}\\ 2:} What are the results of DriveZero on NAVSIM v1 set? The paper discusses DriveZero in the main text, but its results are not included in Table 1.

Response2:\color{blue}{\mathit{Response}\\ 2:} Thank you for pointing this out. For fair comparison, Sec. 5.2 uses the result for DriveZero with DriveVFM ViT-S. This variant uses NAVSIM navtrain without SimScale for downstream training (DriveZero, arXiv:2609.06055v1, Tabs. 7 and A13).

MethodNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
DriveZero99.299.196.310091.894.4
ConDenseDrive (ours)99.398.896.810091.994.5

ConDenseDrive achieves better performance despite using a narrower distillation corpus and no RL-generated trajectory supervision. Our design changes are confined to the visual encoder, leaving the downstream planner architecture unchanged.

AuthorsOpenReview ↗

Anticipating Your Participation as Reviewer-Author Discussion Deadline Approaching

Thank you for your time and thoughtful feedback. We have provided comprehensive responses to all the concerns. As the discussion period ends in two days, we would appreciate knowing whether any questions remain. We are happy to provide further clarification. Thank you again for your time and consideration.

The authors have addressed my concerns, and I will keep my original score.

Reviewer_6bypOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ConDenseDrive, a three-stage framework for improving visual representations and inference efficiency in end-to-end autonomous driving. It first distills knowledge from multiple vision foundation models into a compact encoder, then introduces multi-frame register adaptation and register-guided token compression. Experiments on NAVSIM v1, v2, and HUGSIM show promising planning performance and improved inference efficiency, and a 2X speedup through token compression.

Critical strengths

S1. The idea of improving the visual encoder rather than redesigning the downstream planner is interesting. Combining knowledge distillation with token compression is a practical direction, especially for autonomous driving where inference efficiency is important. The results show that a relatively small encoder can achieve strong planning performance while reducing computation.

S2. The experiments are fairly comprehensive, covering NAVSIM v1 and v2, and zero-shot closed-loop evaluation on HUGSIM. I also see the accuracy-efficiency analysis and the ablations across the three stages. The 2X speedup without sacrificing planning performance is particularly interesting. The additional experiments in the appendix are also helpful, especially the transfer to DiffusionDrive and the analysis of temporal attention.

Critical weaknesses

W1. The advantage of the register-guided token compression is not so clear. As discussed in Section 2, EViT already uses attention scores to select informative tokens and fuse the remaining ones. The main difference here seems to be using register tokens instead of CLS tokens. However, Table 5 only shows a 0.3 PDMS improvement.

W2. The ablation study may not be sufficient. Table 4 shows that most of the performance improvement is from multi-teacher distillation, but I did not find an ablation showing what each teacher contributes to the final planning performance. Appendix C provides additional experiments on temporal attention, which are helpful. However, the improvement from multi-frame modeling is still relatively small compared with the additional computation. I would like to better understand whether this additional complexity is really needed and how much each component contributes to the final results.

Questions for authors

Q1. How much does each teacher contribute to the final planning performance? Have you compared different teacher combinations?

Q2. How does register-guided token compression compare with existing attention-based token compression methods under the same setting?

Q3. I would like to better understand the advantages and limitations of ConDenseDrive compared with VLM-based and diffusion-based driving methods, especially in terms of planning performance and inference efficiency. Appendix B shows that the distilled encoder improves DiffusionDrive, but does the full framework with the token compression also work with different planning architectures?

Q4. Are the reported 94.5 PDMS on NAVSIM v1 and 90.8 EPDMS on NAVSIM v2 evaluated using exactly the same checkpoint, or are the checkpoints trained or selected separately?

Q5. Why does the HUGSIM evaluation use T = 1 and ρ = 0.7 instead of the selected T = 2 and ρ = 0.4 configuration? Is there a general rule for chosing these hyperparameters?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 6byp 1/2

We thank the reviewer for the thoughtful and constructive feedback. Our responses to each point are provided below.

Question1:\color{red}{\mathit{Question}\\ 1:} How much does each teacher contribute to the final planning performance? Have you compared different teacher combinations?

Response1:\color{blue}{\mathit{Response}\\ 1:} We appreciate the reviewer’s feedback. We additionally compare each teacher individually with the full three-teacher combination on the same DeiT-S student under the same downstream training settings on NAVSIM v1 (T = 1, ρ = 1.0).

EncoderTeacherNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
DeiT-S-98.797.995.610088.091.8
DeiT-SDINOv399.298.896.910090.493.9(+2.1)
DeiT-SSigLIP298.998.696.510089.493.1(+1.3)
DeiT-SDA398.998.296.110087.892.1(+0.3)
DeiT-SDINOv3, SigLIP2, DA399.198.896.610091.294.1(+2.3)

DINOv3 yields the strongest single-teacher result at 93.9 PDMS, followed by SigLIP2 (93.1) and DA3 (92.1). The three-teacher model achieves the highest PDMS (94.1), outperforming every single-teacher variant and improving on undistilled DeiT-S by 2.3 points. These results suggest that the shared student representation captures complementary appearance and geometric cues from multiple teachers.

Question2:\color{red}{\mathit{Question}\\ 2:} The advantage of the register-guided token compression is not so clear. How does register-guided token compression compare with existing attention-based token compression methods under the same setting?

Response2:\color{blue}{\mathit{Response}\\ 2:} Thank you for this question. In the single-frame ablation, replacing CLS attention with register attention increases PDMS from 94.0 to 94.3 (Tab. 5). To test whether this gain extends to two-frame inputs, we compare the two routing signals at ρ = 0.4 and 0.7.

FramesKeep rateRouting tokenNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑FPS ↑
T = 2ρ = 0.4CLS98.898.996.410091.294.027.3
T = 2ρ = 0.4Register99.398.896.810091.994.5(+0.5)26.6
T = 2ρ = 0.7CLS98.998.996.110091.794.122.0
T = 2ρ = 0.7Register99.198.996.210092.394.4(+0.3)20.9

At T = 2, register routing improves PDMS over CLS routing by 0.5 points at ρ = 0.4 and by 0.3 points at ρ = 0.7. These results demonstrate that routing with driving-oriented registers achieves better planning performance than CLS routing under the same setting.

Question3:\color{red}{\mathit{Question}\\ 3:} I would like to better understand whether this additional complexity is really needed and how much each component contributes to the final results.

Response3:\color{blue}{\mathit{Response}\\ 3:} Thank you for raising this point. Tab. 4 shows that distillation provides the largest gain, from 91.8 to 94.1 PDMS, without changing throughput. Adding a second frame with temporal attention gives a smaller gain, from 94.1 to 94.3 PDMS, while reducing throughput from 25.4 to 13.6 FPS. The ablation in Appendix C shows that, with T fixed at 2, enabling temporal attention increases PDMS from 94.0 to 94.3. Register-guided compression then raises the two-frame model to 94.5 PDMS and restores throughput to 26.6 FPS.

We use two frames in the main configuration because, with compression, the model gains 0.4 PDMS over the dense single-frame model without sacrificing throughput.

AuthorsOpenReview ↗

Response to Reviewer 6byp 2/2

Question4:\color{red}{\mathit{Question}\\ 4:} I would like to better understand the advantages and limitations of ConDenseDrive compared with VLM-based and diffusion-based driving methods, especially in terms of planning performance and inference efficiency.

Response4:\color{blue}{\mathit{Response}\\ 4:} ConDenseDrive improves planning performance and inference efficiency through a visual encoder design that can be adapted to different downstream planners. We use DiffusionDriveV2 (arXiv:2512.07745v1) and ReCogDrive (ICLR 2026) as representative baselines for diffusion-based and VLM-based driving, respectively. To assess the trade-off between planning performance and inference efficiency, we compare ConDenseDrive with both baselines on NAVSIM v1 using PDMS and end-to-end throughput (FPS).

MethodNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑FPS ↑
DiffusionDriveV2 (diffusion-based)98.397.994.899.987.591.240.0
ReCogDrive (VLM-based)97.997.394.910087.390.83.8
ConDenseDrive (T = 2, ρ = 0.4)99.398.896.810091.994.526.6
ConDenseDrive (T = 1, ρ = 0.3)99.198.796.610091.094.044.5

Our single-frame model outperforms both in planning performance and inference efficiency, while our two-frame model achieves the highest PDMS. In both configurations, our architectural changes are confined to the visual encoder, leaving the downstream planner unchanged. Compared with diffusion-based and VLM-based driving methods, ConDenseDrive achieves a better trade-off between planning performance and inference efficiency.

Question5:\color{red}{\mathit{Question}\\ 5:} Appendix B shows that the distilled encoder improves DiffusionDrive, but does the full framework with the token compression also work with different planning architectures?

Response5:\color{blue}{\mathit{Response}\\ 5:} Thank you for raising this important point. ConDenseDrive confines its design changes to the visual encoder, combining multi-teacher distillation, temporal modeling, and patch-token compression. The full framework can extend beyond DrivoR to other register-based planners, including DriveZero (arXiv:2609.06055), RRDrive (arXiv:2609.21486), and MORE (arXiv:2610.14871). Compression operates within the visual encoder, so the downstream planner can retain its register interface.

Question6:\color{red}{\mathit{Question}\\ 6:} Are the reported 94.5 PDMS on NAVSIM v1 and 90.8 EPDMS on NAVSIM v2 evaluated using exactly the same checkpoint, or are the checkpoints trained or selected separately?

Response6:\color{blue}{\mathit{Response}\\ 6:} Thank you for pointing this out. The NAVSIM v1 and v2 results (94.5 PDMS and 90.8 EPDMS, respectively) come from the same checkpoint. We did not train or select separate checkpoints for the two benchmarks.

Question7:\color{red}{\mathit{Question}\\ 7:} Why does the HUGSIM evaluation use T = 1 and ρ = 0.7 instead of the selected T = 2 and ρ = 0.4 configuration? Is there a general rule for chosing these hyperparameters?

Response7:\color{blue}{\mathit{Response}\\ 7:} Thank you for pointing this out. We evaluated both configurations on the same 436 scenarios under the same protocol, without HUGSIM training or fine-tuning. We report the higher-scoring configuration, T = 1 and ρ = 0.7.

FramesKeep rateHD-Score ↑
T = 1ρ = 0.737.4
T = 2ρ = 0.431.9
Reviewer_Bsk1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes ConDenseDrive, a three-stage framework combining multi-teacher visual distillation, multi-frame register adaptation, and register-guided token compression for end-to-end autonomous driving. Evaluations on NAVSIM v1/v2 and HUGSIM demonstrate strong planning performance with a compact encoder and reduced inference latency.

Critical strengths
  1. The motivation is clear: improving visual representations while reducing the computational cost of the encoder addresses a practical bottleneck.

  2. The measured inference speedup, with little change in accuracy across compression ratios, is promising for deployment.

Critical weaknesses
  1. The contribution relative to closely related work could be better established. ConDenseDrive shares its multi-teacher distillation and DrivoR-style planning design with DriveZero, while its compression resembles EViT-style ranking and fusion. Temporal adaptation and register-guided routing are useful extensions, and the efficiency gains are encouraging. However, the routing comparison in Table 5 covers only one setting. The available evidence leaves uncertainty about the benefit specific to the proposed design, which limits my assessment of its methodological contribution.

  2. The main accuracy improvement is not sufficiently isolated. The +2.3 PDMS from distillation in Table 4 is measured against the original DeiT-S initialization. Controlled comparisons with stronger visual encoders and single-teacher distillation are missing. It is therefore unclear how much improvement comes from combining complementary teachers rather than stronger pretraining or driving-domain adaptation. Since distillation accounts for most of the reported accuracy gain, this uncertainty materially affects my assessment of the central contribution.

Questions for authors
  1. Could you clarify how ConDenseDrive differs from DriveVFM/DriveZero in its visual distillation and planner adaptation? It would also be helpful to identify the DriveZero configuration and score referenced in the PDMS comparison in Section 5.2.

  2. Do you have comparisons with DINOv2-S or DINOv3-S initialization and single-teacher distillation under the same downstream settings? These would help determine whether combining teachers provides a distinct planning benefit.

  3. Does the advantage of register-guided routing over CLS routing hold at other compression ratios, particularly the main two-frame configuration? A comparison with an existing compression method at matched inference cost would help establish whether the proposed routing improves the accuracy–efficiency trade-off.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Bsk1 1/2

We thank the reviewer for the valuable suggestions. We appreciate the positive assessment and address the comments below.

Question1:\color{red}{\mathit{Question}\\ 1:} Could you clarify how ConDenseDrive differs from DriveVFM/DriveZero in its visual distillation and planner adaptation?

Response1:\color{blue}{\mathit{Response}\\ 1:} Thank you for raising this comparison. We note that DriveZero (arXiv:2609.06055v1) is a concurrent preprint. The two methods differ in their visual distillation, planner adaptation, and visual encoding.

(1) Visual distillation.

AspectDriveVFMConDenseDrive
TeachersDINOv3, SigLIP2, SAM, Depth Anything V2DINOv3, SigLIP2, Depth Anything 3
SupervisionCosine loss on summary tokens; MSE on patch tokensCosine and Smooth L1 losses on patch tokens
Distillation dataLAION-2B, ImageNet-21K, SA-1B, OpenDV, nuPlan, WaymoNAVSIM navtrain

The two methods use different teacher sets. DriveVFM supervises both summary tokens and patch features, whereas ConDenseDrive distills patch features alone. Despite using a narrower distillation corpus, ConDenseDrive achieves comparable planning performance on NAVSIM v1.

(2) Planner adaptation. Both methods use a DrivoR-style planner. However, DriveZero supervises trajectory generation with DriveRL rollouts and goal augmentation. ConDenseDrive retains DrivoR’s planner architecture and training objectives without RL-generated trajectory supervision. With downstream planner architecture unchanged, ConDenseDrive achieves competitive planning performance and a 2.0× end-to-end speedup by optimizing visual encoder .

(3) Visual encoding. DriveZero uses current-frame images, whereas ConDenseDrive integrates temporal context through multi-frame register adaptation and compresses patch tokens within the encoder. The full framework can be adapted to register-based driving systems beyond DrivoR, including DriveZero, RRDrive (arXiv:2609.21486), and MORE (arXiv:2610.14871), by adapting the encoder to their register interfaces without redesigning the downstream planner.

Question2:\color{red}{\mathit{Question}\\ 2:} It would also be helpful to identify the DriveZero configuration and score referenced in the PDMS comparison in Section 5.2.

Response2:\color{blue}{\mathit{Response}\\ 2:} Thank you for pointing this out. For fair comparison, Sec. 5.2 uses the result for DriveZero with DriveVFM ViT-S. This variant uses NAVSIM navtrain without SimScale for downstream training (DriveZero, arXiv:2609.06055v1, Tabs. 7 and A13) .

MethodNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
DriveZero99.299.196.310091.894.4
ConDenseDrive (ours)99.398.896.810091.994.5

Although DriveZero uses a broader distillation corpus and trains its planner with DriveRL rollouts and goal augmentation, ConDenseDrive achieves a slightly higher PDMS by optimizing the visual encoder without redesigning the downstream planner.

AuthorsOpenReview ↗

Response to Reviewer Bsk1 2/2

Question3:\color{red}{\mathit{Question}\\ 3:} Do you have comparisons with DINOv2-S or DINOv3-S initialization and single-teacher distillation under the same downstream settings?

Response3:\color{blue}{\mathit{Response}\\ 3:} Thank you for suggesting these comparisons. We evaluate direct DINOv3-S initialization and single- and three-teacher distillation of DeiT-S under the same downstream training settings on NAVSIM v1 (T = 1, ρ = 1.0). We include the published DINOv2-S result from DrivoR (CVPR 2026) as an external reference.

EncoderTeacherNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
DINOv2-S (DrivoR)-99.098.996.710090.093.7
DeiT-S (baseline)-98.797.995.610088.091.8
DINOv3-S-99.098.596.310089.493.0(+1.2)
DeiT-SDINOv399.298.896.910090.493.9(+2.1)
DeiT-SSigLIP298.998.696.510089.493.1(+1.3)
DeiT-SDA398.998.296.110087.892.1(+0.3)
DeiT-SDINOv3, SigLIP2, DA399.198.896.610091.294.1(+2.3)

DeiT-S distilled from DINOv3 ViT-L/16 alone achieves 93.9 PDMS, compared with 93.0 for direct DINOv3-S initialization. Among the single-teacher variants, DINOv3 performs best, followed by SigLIP2 at 93.1 PDMS and DA3 at 92.1 PDMS. Combining all three teachers yields 94.1 PDMS, 2.3 points above undistilled DeiT-S and 0.2 points above the strongest single-teacher variant. These results demonstrate the effectiveness of our multi-teacher visual distillation for downstream planning.

Question4:\color{red}{\mathit{Question}\\ 4:} Does the advantage of register-guided routing over CLS routing hold at other compression ratios, particularly the main two-frame configuration? A comparison with an existing compression method at matched inference cost would help establish whether the proposed routing improves the accuracy–efficiency trade-off.

Response4:\color{blue}{\mathit{Response}\\ 4:} We appreciate the request for a comparison at matched compression budgets. We have additionally compared CLS and register routing in the two-frame configuration at ρ = 0.4 and ρ = 0.7.

FramesKeep rateRouting tokenNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑FPS ↑
T = 2ρ = 0.4CLS98.898.996.410091.294.027.3
T = 2ρ = 0.4Register99.398.896.810091.994.5(+0.5)26.6
T = 2ρ = 0.7CLS98.998.996.110091.794.122.0
T = 2ρ = 0.7Register99.198.996.210092.394.4(+0.3)20.9

At a fixed keep rate, both routing variants retain the same number of tokens at each compression block and achieve comparable end-to-end throughput. Register-guided compression improves PDMS over CLS routing by 0.5 points at ρ = 0.4 and by 0.3 points at ρ = 0.7. Its PDMS varies by only 0.1 across the two keep rates. These results show that our register-guided compression improves planning performance at the same token budget.

AuthorsOpenReview ↗

Anticipating Your Participation as Reviewer-Author Discussion Deadline Approaching

Thank you for your time and thoughtful feedback. We have provided comprehensive responses to all the concerns. As the discussion period ends in two days, we would appreciate knowing whether any questions remain. We are happy to provide further clarification. Thank you again for your time and consideration.

Thank you for your detailed rebuttal and clarifications. I have no further questions.

Comments on the paper

8 comments
AuthorsOpenReview ↗

Response to Reviewer 6byp

Thank you for this follow-up. For a new benchmark, we recommend choosing T first based on planning performance and then selecting ρ at that fixed T to balance planning performance and inference latency.

AuthorsOpenReview ↗

Response to Reviewer 6byp

Thank you for raising this concern. No navtest images were used for multi-teacher visual distillation. The frozen teachers extract features exclusively from front-view images in navtrain.

AuthorsOpenReview ↗

Response to Reviewer 6byp

Thank you for the question. ConDenseDrive uses only the NAVSIM navtrain split for distillation. All three frozen teachers extract features from front-view images in navtrain.

AspectDriveVFMConDenseDrive
TeachersDINOv3, SigLIP2, SAM, Depth Anything V2DINOv3, SigLIP2, Depth Anything 3
Distillation dataLAION-2B, ImageNet-21K, SA-1B, OpenDV, nuPlan, WaymoNAVSIM navtrain

ConDenseDrive achieves comparable planning performance on NAVSIM v1 with less distillation data than DriveZero.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.