acceptodds
Under review as a conference paper at ICLR 2027

UniWAM: Unified World-Action Model

21 views

Abstract

Vision-language-action models benefit from the understanding and reasoning capabilities of pretrained vision-language models, but action-only supervision provides limited grounding in world dynamics. Conversely, world-action models inherit spatiotemporal priors from video generation models, yet remain limited in semantic understanding and reasoning under distribution shifts. We introduce UniWAM, a unified architecture that integrates a physical reasoner, a world generator, and an action predictor to jointly learn semantic understanding of the physical world, visual generation, and action prediction. To adapt the vision-language component to embodied tasks while preserving its inherited language capabilities, we represent low-level robot actions in natural language and introduce a pre-training recipe that assigns complementary supervision from visual question answering (VQA) data, human egocentric data, and robot demonstrations to the appropriate model components. During post-training, future visual noise augmentation reduces reliance on precise future predictions, while history-conditioned flow matching uses encoded action history to initialize action generation. Together, these designs significantly reduce denoising steps while maintaining performance. UniWAM achieves state-of-the-art (SOTA) performance across multiple evaluations, including in-distribution performance, robustness, generalization, instruction following, and long-horizon task execution.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied Foundation Model, Robot Manipulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_1v8mOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UniWAM, a unified World-Action Model integrating physical reasoning, world generation, and action prediction. The main contributions include: (1) A Mixture-of-Transformers architecture that enables information exchange among vision-language reasoning, future video generation, and robotic action prediction through joint multimodal attention. (2) A multi-source pretraining strategy with physical-language supervision, leveraging VQA, human egocentric videos, and robot demonstrations. (3) Future-Frame Noise Augmentation and History-Conditioned Flow Matching to improve action prediction robustness and inference efficiency. Experiments on LIBERO, LIBERO-Plus, RoboTwin 2.0, and real-world robotic tasks demonstrate competitive performance in manipulation success, out-of-distribution generalization, instruction following, and long-horizon execution.

Critical strengths
  1. UniWAM integrates semantic reasoning, world generation, and action prediction into a unified framework with strong practical relevance.
  2. The model achieves competitive results on LIBERO, LIBERO-Plus, and RoboTwin 2.0, particularly demonstrating strong out-of-distribution generalization.
  3. Real-world robotic experiments and multiple ablations support the effectiveness of multi-source pretraining, physical-language supervision, and history-conditioned action generation.
Critical weaknesses
  1. UniWAM shares substantial similarities with Motus in its three-expert MoT architecture and unified multimodal modeling. Although the authors emphasize differences in training recipes, data composition, and model capabilities, the distinct contributions of these improvements over Motus are not sufficiently clarified.
  2. The method relies on a relatively large unified Transformer architecture. Although some action-generation inference times are reported, a systematic analysis of overall computational cost, memory requirements, and scalability under constrained hardware settings is still lacking.
Questions for authors
  1. What are the key differences between UniWAM and Motus in terms of training recipes, data composition, and model capabilities? Can the authors further clarify the independent contributions of the proposed designs?
  2. Can the authors provide additional analysis of computational cost and inference efficiency, including GPU memory consumption, inference latency, and deployment feasibility under limited GPU resources?
  3. It would be interesting to discuss how UniWAM could be extended to more complex 3D environments involving geometric structures, occlusions, and multi-object interactions, and what architectural or training modifications might be required.
4 comments
AuthorsOpenReview ↗

Response to Reviewer 1v8m(1/2)

[W1] UniWAM shares substantial similarities with Motus in its three-expert MoT architecture and unified multimodal modeling. Although the authors emphasize differences in training recipes, data composition, and model capabilities, the distinct contributions of these improvements over Motus are not sufficiently clarified.

Response: Thank you for this comment. We acknowledge that UniWAM builds on the three-expert Mixture-of-Transformers (MoT) architecture of Motus. In the following, we will clarify our improvements over Motus, which constitute our distinct contributions.

1. Training recipe. Whereas Motus uses a frozen VLM as a multimodal encoder, we train the VLM-based physical reasoner during both pretraining and post-training, using physical language supervision to learn a shared mapping between semantics and actions while preserving pretrained language capabilities. As shown in Figure 5(b), unfreezing the VLM and adding physical language supervision progressively improve downstream performance. During post-training, history-conditioned flow matching uses recent action history as an initialization prior, improving action prediction performance and reducing the required denoising steps. Future-frame noise augmentation encourages the action predictor to tolerate imperfect future visual predictions, providing additional performance gains (see Figure 5(c)).

2. Data composition. We combine human egocentric data, robot teleoperation data, and visual question answering (VQA) data. In contrast, Motus includes neither VQA data nor action labels for its human data. We apply physical language supervision to both human and robot actions, while VQA data provides complementary visual–semantic and embodied reasoning supervision. Figure 5(a) shows that the three-source mixture achieves the highest average success rate among the evaluated data combinations. Human–robot co-training also yields increasing benefits at larger data scales (see Figure 4(a)).

3. Model capabilities. These designs support instruction following and generalization across environments. As shown in Figure 3(a–b), UniWAM achieves strong real-world Instruction-Following Rate and Task Success Rate compared with Motus and other baselines. Its generalization is further evaluated on RoboTwin 2.0 clean-to-random (C2R) and LIBERO-Plus (see Tables 1 and 2). Together with the component ablations, these results support the benefits of our training recipe and data composition.

We will make these distinctions and their supporting evidence more explicit in the camera-ready version.

[W2] The method relies on a relatively large unified Transformer architecture. Although some action-generation inference times are reported, a systematic analysis of overall computational cost, memory requirements, and scalability under constrained hardware settings is still lacking.

Response: Please see our response to Q2.

[Q1] What are the key differences between UniWAM and Motus in terms of training recipes, data composition, and model capabilities? Can the authors further clarify the independent contributions of the proposed designs?

Response: Please see our response to W1.

[Q2] Can the authors provide additional analysis of computational cost and inference efficiency, including GPU memory consumption, inference latency, and deployment feasibility under limited GPU resources?

Response: We have evaluated UniWAM's inference cost on NVIDIA H800, RTX 4090, and Jetson Thor T5000. All evaluations use four denoising steps and generate 16 actions per chunk, with an estimated computational cost of approximately 17,200 GFLOPs per chunk.

HardwareEstimated GFLOPs / chunkPeak GPU memory (GB)Latency / chunk (ms)Action frequency (Hz)
NVIDIA H800~17,20018.962925.44
NVIDIA RTX 4090~17,20019.11,47710.83
NVIDIA Jetson Thor T5000~17,20018.62,0167.94

Chunk latency measures the end-to-end time from receiving image observations to outputting the complete action chunk.

For Jetson Thor, GPU memory usage is drawn from unified system memory (128 GB).

As shown in the table above, UniWAM runs on all three platforms with peak GPU memory usage below 20 GB. The RTX 4090 and Jetson Thor T5000 achieve action-generation throughputs of 10.83 and 7.94 actions/s, respectively, demonstrating deployment feasibility on consumer and embedded hardware. We will explore asynchronous inference to overlap action generation with execution and reduce deployment idle time in future work.

AuthorsOpenReview ↗

Response to Reviewer 1v8m(2/2)

[Q3] It would be interesting to discuss how UniWAM could be extended to more complex 3D environments involving geometric structures, occlusions, and multi-object interactions, and what architectural or training modifications might be required.

Response: Thank you for this suggestion. We consider three complementary directions for extending UniWAM to more complex 3D environments.

1. Geometric structures. 1) Implicit alignment, as in Spatial Forcing [1], could be incorporated by aligning UniWAM's internal representations with geometry-aware features during training. 2) Explicit 3D inputs, as in SpatialVLA [2], could be incorporated by adding depth-derived 3D positional encodings to the visual tokens of our VLM-based physical reasoner. We have conducted preliminary experiments incorporating Spatial Forcing during post-training by aligning intermediate-layer features of the VLM-based physical reasoner with geometry-aware features from VGGT. The results are summarized below (C2C: clean-to-clean; C2R: clean-to-random).

ConfigurationRoboTwin 2.0 C2C (%)RoboTwin 2.0 C2R (%)Average (%)
UniWAM75.1468.3271.73
UniWAM + Spatial Forcing78.7270.4174.57

As shown in the table above, incorporating implicit spatial representation alignment inspired by Spatial Forcing improves both C2C and C2R success, increasing average success from 71.73% to 74.57%. These preliminary results support geometric representation alignment as a promising extension to UniWAM. We leave the evaluation of explicit 3D inputs to future work.

2. Occlusions. SaPaVe's active-perception strategy [3] offers a feasible direction for extending UniWAM: the VLM-based physical reasoner could guide viewpoint selection, while a separate camera-pose prediction head could generate camera movements alongside the existing manipulation actions. This would require training the added head on image–language–camera movement data, followed by joint training on camera-control and manipulation trajectories with changing viewpoints. With a movable camera, the predicted movements could reveal occluded objects and provide updated observations for subsequent action prediction.

3. Multi-object interactions. We could extend UniWAM in two ways: by incorporating more multi-object interaction data during training, and by introducing an agent as a high-level planner to specify the order of object interactions and the execution plan. We have conducted preliminary experiments integrating GPT-6 Astra as the high-level planner, which passes its execution plans to UniWAM as text instructions for low-level action execution. We evaluate this extension on selected multi-object interaction tasks from RoboTwin 2.0, with results summarized below.

Configurationstack_blocks_three (C2C/C2R, %)put_object_cabinet (C2C/C2R, %)blocks_ranking_size (C2C/C2R, %)
UniWAM74.00 / 52.0040.00 / 28.0036.00 / 38.00
UniWAM + GPT-6 Astra high-level planner76.00 / 59.0047.00 / 39.0048.00 / 46.00

As shown in the table above, adding the high-level planner improves success on all three tasks under both C2C and C2R, with C2R gains of 7, 11, and 8 percentage points, respectively. These results suggest that high-level planning can complement UniWAM's action prediction for multi-object interactions.

We will expand the discussion of these extensions and future research directions in the camera-ready version.

References

[1] Li, F., et al. Spatial Forcing: Implicit Spatial Representation Alignment for Vision-language-action Model. ICLR, 2026. https://openreview.net/forum?id=euMVC1DO4k

[2] Qu, D., et al. SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Model. RSS, 2025. https://arxiv.org/abs/2501.15830

[3] Liu, M., et al. SaPaVe: Towards Active Perception and Manipulation in Vision-Language Action Models for Robotics. CVPR, 2026. https://openaccess.thecvf.com/content/CVPR2026/papers/Liu_SaPaVe_Towards_Active_Perception_and_Manipulation_in_Vision-Language_Action_Models_CVPR_2026_paper.pdf

Dear Reviewer 1v8m,

I hope this message finds you well. As the discussion period is nearing its end with less than five days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and were eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Best regards, Authors.

Thank you for the detailed rebuttal and clarifications. I appreciate the authors' efforts in addressing my concerns. The paper presents a unified framework with promising experimental results and practical relevance. I will maintain my original score of 3.

Reviewer_tC5oOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces UniWAM, which is jointly trained on VQA data, human egocentric data, and robot data. By injecting noise, it reduces reliance on precise future predictions. UniWAM performs well in both IID and OOD settings.

Critical strengths
  1. The figures and tables are visually appealing and clearly presented.
  2. The method is well motivated.
Critical weaknesses
  1. The noise augmentation strategy resembles previous methods to mitigating reliance on teacher forcing, limiting its novelty.
  2. This paper lacks theoretical analysis about the method's effect on generalization ability.
Questions for authors
  1. Could the authors provide a theoretical analysis of the mechanisms underlying the improved generalization?
  2. How does the proposed noise augmentation differ from representative previous methods to mitigating reliance on teacher forcing?
  3. Could the authors further analyze how the data mixture, particularly the proportions of different data sources, affects model performance?
  4. Could the authors provide more comprehensive ablation studies to isolate the contributions of individual components, such as the model architecture, history-conditioned flow matching, and noise augmentation?
5 comments
AuthorsOpenReview ↗

Response to Reviewer tC5o(1/2)

[W1] The noise augmentation strategy resembles previous methods to mitigating reliance on teacher forcing, limiting its novelty.

Response: We acknowledge that future-frame noise augmentation itself is not a core novelty of our work. We will clarify this distinction in the camera-ready version and more explicitly identify the contribution of our post-training design.

Our main contribution in post-training is history-conditioned flow matching, which uses recent action history as a prior for action initialization. Figure 5(c) in the main paper shows that this design increases the average success rate from 64.80% to 69.73% compared with noise-to-action generation, a gain of 4.93 percentage points.

Furthermore, to the best of our knowledge, we are the first to investigate the synergy between history-conditioned flow matching and future-frame noise augmentation within a unified world-action model. As shown in Figure 5(c), adding future-frame noise augmentation further increases the average success rate from 69.73% to 71.73%, a gain of 2.00 percentage points over history-conditioned flow matching alone. This supports the complementary roles of the two strategies.

[W2] This paper lacks theoretical analysis about the method's effect on generalization ability.

Response: Thank you for this comment. We provide a theoretical analysis of the method's effect on generalization ability from two perspectives.

First, human egocentric data exposes the model to diverse scenes and manipulation tasks beyond robot demonstrations. This broader coverage can reduce reliance on environment-specific visual cues and support transfer to unfamiliar settings.

Second, physical language supervision provides a shared semantic representation for human and robot actions. By expressing actions from different embodiments in a common language format, it encourages the model to learn transferable manipulation semantics rather than embodiment-specific correlations.

Together, data diversity broadens the model’s experience, while shared physical language supervision facilitates transfer across embodiments. This explanation is consistent with the human–robot co-training analysis in Figure 4(a–b), the ablation results in Figure 5(a–b), and the generalization results in Tables 1 and 2. We will include this discussion in the camera-ready version.

[Q1] Could the authors provide a theoretical analysis of the mechanisms underlying the improved generalization?

Response: Please see our response to W2.

[Q2] How does the proposed noise augmentation differ from representative previous methods to mitigating reliance on teacher forcing?

Response: Please see our response to W1.

AuthorsOpenReview ↗

Response to Reviewer tC5o(2/2)

[Q3] Could the authors further analyze how the data mixture, particularly the proportions of different data sources, affects model performance?

Response: Thank you for this suggestion. We evaluated three additional data mixtures on RoboTwin 2.0. C2C: clean-to-clean; C2R: clean-to-random.

No.RobotHumanVQARoboTwin 2.0 C2C (%)RoboTwin 2.0 C2R (%)Average (%)
1 (ours)0.400.500.1075.1468.3271.73
20.600.300.1076.2364.6970.46
30.200.700.1073.1568.0770.61
40.360.450.1972.3263.2867.80

Human–robot balance. Comparing Rows 1–3, a higher robot proportion favors C2C performance, whereas retaining more human data supports stronger C2R performance. However, increasing the human proportion to 0.70 yields no further C2R improvement and lowers C2C success. Our default mixture achieves the best average performance among these settings.

VQA proportion. Comparing Rows 1 and 4, increasing the VQA proportion while preserving the robot–human ratio reduces average success from 71.73% to 67.80%. This suggests that a larger share of VQA data cannot substitute for action data in supporting manipulation performance.

These results support our default Robot/Human/VQA sampling proportions of 0.40/0.50/0.10. We will clarify this configuration and include the analysis in the camera-ready version.

[Q4] Could the authors provide more comprehensive ablation studies to isolate the contributions of individual components, such as the model architecture, history-conditioned flow matching, and noise augmentation?

Response: Thank you for this suggestion. Figure 5(c) already evaluates history-conditioned flow matching and its combination with future-frame noise augmentation. We extend this analysis by adding an Future-frame noise augmentation variant and evaluating all four configurations on Cosmos3-Edge.

ArchitectureHistory-conditioned flow matchingFuture-frame noise augmentationC2C ↑C2R ↑Average ↑
ours✗✗69.2460.3564.80
ours✓✗74.2365.2269.73
ours✗✓70.8264.6867.75
ours✓✓75.1468.3271.73
Cosmos3-Edge✗✗32.1816.7924.49
Cosmos3-Edge✓✗36.6818.8427.76
Cosmos3-Edge✗✓33.1119.2326.17
Cosmos3-Edge✓✓39.4523.3731.41

Individual and joint contributions. On our architecture, history-conditioned flow matching and future-frame noise augmentation respectively increase average success from 64.80% to 69.73% and 67.75%. Using both achieves 71.73%, outperforming either component alone and supporting their complementary roles.

Transfer across architectures. The same trend holds on Cosmos3-Edge, where the two components individually improve average success from 24.49% to 27.76% and 26.17%, and jointly achieve 31.41%. Unlike our model, Cosmos3-Edge lacks pretraining on dual-arm manipulation data, which may partly explain its lower overall performance on RoboTwin 2.0. Although absolute performance differs between architectures, both components consistently provide gains, supporting their applicability beyond our architecture.

We will present these ablations more systematically in the camera-ready version.

Dear Reviewer tC5o,

I hope this message finds you well. As the discussion period is nearing its end with less than five days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and were eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Best regards, Authors.

Reviewer_yE8WOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

UniWAM combines a language reasoner, video world model, and action predictor in a three-expert Mixture-of-Transformers architecture. Physical language supervision connects robot actions to the language stream. The paper also introduces a multi-source training recipe, future-visual noise augmentation, and history-conditioned action generation, evaluated in simulation and on real robots.

Critical strengths
  1. The paper presents a useful integration of language reasoning, future prediction, and control. Evaluations across simulation, real robots, and distribution shifts support the effectiveness of the combined system.

  2. The data-source and action-generation ablations make the training recipe informative. The human–robot co-training study also provides useful evidence about when adding human data helps.

Critical weaknesses
  1. The architectural novelty is limited. The three-expert MoT design is closely related to Motus, while physical language supervision follows LAP. Both are acknowledged. The contribution is therefore primarily the training recipe and its empirical effectiveness, which should be reflected more clearly in the claims.

  2. Preserving pretrained language capabilities is not directly established. Better manipulation instruction following and overlapping feature projections do not measure whether general language or VQA performance is retained after training.

Questions for authors
  1. Can the authors report held-out language or VQA performance before and after training to support the preservation claim?

  2. How much does physical language supervision contribute under the full pretraining setup when data and optimization budgets are held fixed? The cumulative ablation does not fully isolate this effect at scale.

4 comments
AuthorsOpenReview ↗

Response to Reviewer yE8W(1/2)

[W1] The architectural novelty is limited. The three-expert MoT design is closely related to Motus, while physical language supervision follows LAP. Both are acknowledged. The contribution is therefore primarily the training recipe and its empirical effectiveness, which should be reflected more clearly in the claims.

Response: We agree that our contribution statement should be rephrased to more clearly emphasize our training recipe, data composition, and model capabilities. We will revise the contribution statement in the camera-ready version as follows:

“Overall, at the architectural level, we build on the MoT architecture of Motus to connect language reasoning, video generation, and action prediction through joint attention. Our contributions are summarized in three aspects.

  • Pre-training. We construct a three-source pre-training scheme comprising human data, robot teleoperation data, and VQA data, and systematically evaluate the contribution of each data source through source-level ablation studies.
  • Post-training. We explicitly formulate how recent action history and perturbed future visual latents are incorporated into the joint flow-matching process.
  • Empirical performance. Extensive evaluations demonstrate that UniWAM achieves state-of-the-art performance across multiple dimensions, including in-distribution performance, robustness, generalization, instruction following, and long-horizon task execution.”

While our method builds on Motus and LAP, there are several key differences and improvements. We clarify them as follows:

Compared with Motus:

  1. Training recipe. We unfreeze the VLM, whereas Motus uses it as a frozen multimodal encoder. We convert human and robot action labels into physical language to learn visual-language-to-action mappings while supporting language understanding. During post-training, we incorporate history-conditioned flow matching and future-frame noise augmentation to improve performance and efficiency.
  2. Data composition. We combine general-purpose VQA data with action-labeled human and robot data to preserve language understanding and enable joint human–robot training. Motus includes neither VQA data nor action labels for its human data.
  3. Model capabilities. These designs improve language understanding and instruction following, and support stronger generalization, as demonstrated on RoboTwin 2.0 clean-to-random and LIBERO-Plus.

Compared with LAP:

  1. We extend physical language supervision to human action data, broadening its applicability to joint human–robot learning.
  2. We integrate this supervision into a unified world-action architecture that combines semantic understanding with world modeling.

[W2] Preserving pretrained language capabilities is not directly established. Better manipulation instruction following and overlapping feature projections do not measure whether general language or VQA performance is retained after training.

Response: To directly assess the retention of pretrained capabilities, we evaluate the VLM component of UniWAM before and after training, using general and embodied VQA benchmarks as representative measures of visual-language understanding.

CategoryBenchmarkBefore training (%)After training (%)
General VQAMMStar52.4748.93
General VQARealWorldQA64.9762.04
General VQACV-Bench73.7578.85
Embodied VQARoboSpatial-Home41.3549.81
Embodied VQAERQA36.5038.00

As shown in the table, general VQA capabilities are largely preserved after UniWAM training. CV-Bench improves by 5.10 percentage points, while MMStar and RealWorldQA show modest decreases of 3.54 and 2.93 percentage points, respectively. These results suggest that UniWAM training largely retains pretrained general visual-language understanding rather than causing uniform degradation. Meanwhile, embodied VQA performance improves on both RoboSpatial-Home (+8.46 pp) and ERQA (+1.50 pp), indicating that UniWAM training further strengthens embodied visual-language understanding.

These results provide direct quantitative evidence for the retention of general VQA capabilities and the improvement of embodied VQA performance. We will include these evaluations and clarify the scope of our preservation claim in the camera-ready version.

AuthorsOpenReview ↗

Response to Reviewer yE8W(2/2)

[Q1] Can the authors report held-out language or VQA performance before and after training to support the preservation claim?

Response: Please see our response to W2.

[Q2] How much does physical language supervision contribute under the full pretraining setup when data and optimization budgets are held fixed? The cumulative ablation does not fully isolate this effect at scale.

Response: Thank you for this suggestion. We agree that the cumulative ablation alone does not isolate the contribution of physical language supervision under pretraining. We therefore add a pretrained variant without physical language supervision to the original ablation. For this controlled comparison, we keep the data, optimization budget, and post-training settings fixed. The only difference is whether physical language supervision is used during pretraining.

The table below combines the original results with this additional variant. C2C and C2R denote clean-to-clean and clean-to-random success rates (%), respectively; Average is their arithmetic mean.

ConfigurationC2CC2RAverage
Frozen VLM52.4032.8242.61
Unfrozen VLM57.2335.7446.49
Unfrozen VLM + Physical language supervision58.6444.8351.74
Unfrozen VLM + Pretraining (new)60.2149.3254.77
Unfrozen VLM + Pretraining + Physical language supervision75.1468.3271.73

Comparing the last two rows, physical language supervision increases average success from 54.77% to 71.73% under matched data and optimization budgets. This additional ablation isolates the effect of physical language supervision at scale, which shows that incorporating physical language supervision during pretraining helps preserve language understanding and improve downstream action prediction.

Dear Reviewer yE8W,

I hope this message finds you well. As the discussion period is nearing its end with less than five days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and were eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Best regards, Authors.

Thank you for the detailed rebuttal and the revisions. The paper now credits the architecture to Motus and physical language supervision to LAP, which resolves my novelty concern. I will maintain my initial rating as Weak Accept.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.