MoVie: Visual Model-Based Policy Adaptation for View Generalization

Visual Reinforcement Learning (RL) agents trained on limited views face significant challenges in generalizing their learned abilities to unseen views. This inherent difficulty is known as the problem of $\textit{view generalization}$. In this work, we systematically categorize this fundamental problem into four distinct and highly challenging scenarios that closely resemble real-world situations. Subsequently, we propose a straightforward yet effective approach to enable successful adaptation of visual $\textbf{Mo}$del-based policies for $\textbf{Vie}$w generalization ($\textbf{MoVie}$) during test time, without any need for explicit reward signals and any modification during training time. Our method demonstrates substantial advancements across all four scenarios encompassing a total of $\textbf{18}$ tasks sourced from DMControl, xArm, and Adroit, with a relative improvement of $\mathbf{33}$%, $\mathbf{86}$%, and $\mathbf{152}$% respectively. The superior results highlight the immense potential of our approach for real-world robotics applications. Videos are available at https://yangsizhe.github.io/MoVie/ .

Paper

Similar papers

Peer review

Reviewer BF5c4/10 · confidence 3/52023-06-17

Summary

This work presents an approach to train model-based RL methods such that it generalize to novel views on multiple RL benchmarks. The method leverages classical STN and frozen encoder to benefit the generalization performance.

Strengths

The method is sound and simple with barely hyperparameters tuning The design choices study such as different moving views and the insight are helpful for research community.

Weaknesses

The writing can be largely improved on section 2 and 3. It is unclear what the model-based RL problem is, and how to define the view generalization, how to map the model to actual control/policy for evaluations, for audience that are not familiar with this line of work. The problem is specific and the method novelty are limited. Comparisons with baselines are unclear. At least some comparison with model-free RL and recent work [1] would be helpful. [1] Multi-View Masked World Models for Visual Robotic Manipulation, Seo et al., 2023

Questions

In the experiments, does the baseline methods see all various views in the training sets as well? I.e. is the baseline method in figure 3 also seeing different views? There are many other choices to solve this problem. Is contrastive learning / or data augmentations on different views also a way to map different views to the same representations? I am not sure the application on real-world robot experiments, although some visualizations are presented.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

See above.

Authorsrebuttal2023-08-19

Thank you for the review and awaiting your response

We sincerely thank you for your efforts in reviewing our paper and the suggestions again. We believe that we have resolved all the concerns mentioned in the review. Should there be any additional concerns, we are more than happy to address them! Thank you very much!

Reviewer E44Y6/10 · confidence 3/52023-07-06

Summary

This paper mainly provides a training paradigm. Utilizing the dynamic transition model of the environment during the testing phase as a supervisory signal, STN is used to quickly finetune the mapping of observed potential states, resulting in better performance of the strategy mapping trained on a single view task on unseen test views. This work conducted comparative experiments on three types of tasks and four challenging views, achieving results that surpass existing methods. Ablation experiments were conducted to prove the effectiveness of each module setting.

Strengths

The training paradigm provided in this article can effectively handle the generalization problem when view changes occur, and a thorough comparative experiment and ablation experiment have been conducted on the proposed model. The core idea of reconstructing mapping h using the environmental dynamic transfer model in the paper is effective.

Weaknesses

1. The subscripts for o and a in Formula 2 are missing. 2. There are issues with the baseline selection of IDM+STN. Cheetah run alone cannot prove that IDM+STN is generally better than IDM. It is not a problem to conduct ablation experiments on the Cheetah run model alone, but this cannot be used as sufficient evidence for the superiority of IDM+STN over IDM and thus selecting it as the baseline. 3. Chapter 3 emphasizes the differences between Formula 2 and Formula 1, on one hand, due to the fixed parameters of Network d, and on the other hand due to the use of SAE in Network h. One of the core methods is to fix the parameters of d, which is intuitive to avoid changes in the target domain of h during the update process, resulting in the failure of the policy pi. However, the ablation experiment showed that the setting of d^* did not achieve a consistently effective effect. 4. The main approach is to use the dynamic transfer model of the environment under the new view as supervision to quickly train the changed h '. Can you obtain a generalized h directly based on o'?

Questions

1. Have the training methods of the three baselines in the comparative experiment also been finetuned? How much data was used during finetuning, and is this amount of data acceptable for practical applications? 2. Can the network of h ^ SAE be directly used in training to achieve learning from multiple perspectives of tasks? Directly generalize from an unseen perspective without finetuning. 3. The reason why Movie is better than IDM+STN should be explained.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The authors have discussed the limitations

Authorsrebuttal2023-08-19

Thank you for the review and awaiting your response

We sincerely thank you for your efforts in reviewing our paper and the suggestions again. We believe that we have resolved all the concerns mentioned in the review. Should there be any additional concerns, we are more than happy to address them! Thank you very much!

Reviewer N1Ux6/10 · confidence 4/52023-07-06

Summary

The paper addresses the novel problem of view generalization in reinforcement learning, where an RL agent is trained on an environment with a fixed view and then evaluated on a test environment having the exact same dynamics but observed from a different perspective. In order to address this issue, the authors introduce an innovative method that permits test-time adaptation to the new view. The authors integrate a learnable spatial transformer network (STN) into the feature extraction component of the agent. This augmented encoder is then fine-tuned at test time to generate features that enable a frozen latent dynamic model to predict future state representations. An essential aspect of the authors' approach is that this adaptation does not necessitate any reward signal, which is a key advantage. Experimental results demonstrate a significant reduction in the generalization gap presented by a new view across several heterogeneous benchmarks.

Strengths

The paper is well-structured and easy to follow, providing clear and detailed explanations that make believe that I could readily reproduce the method based on the descriptions provided. The authors introduce a new problem in the field that could stimulate further research. The proposed method's strength lies in its test-time adaptation capability, which eliminates the need for any reward signal, making it highly applicable to real-world scenarios. This characteristic also suggests that the approach could be integrated into any model-based method, enhancing its universality. Another advantage is the apparent minimal interaction required with the testing environment, promoting efficiency. The method is evaluated across various environments, with the results showcasing impressive improvements compared to non-adaptive methods. The paper's ablation study effectively illustrates the contribution of each component of the method, highlighting the importance of every aspect in achieving the observed results. Additionally, the annex offers excellent visualizations demonstrating the impact of the spatial transformer on the feature map, providing intuitive understanding of the method's mechanics.

Weaknesses

Despite the many strengths of the paper, a few areas could benefit from further development and clarification. 1. It would be beneficial to see a comparison of the performance decrease relative to the original view. The lack of this data makes it challenging to truly gauge the significance of the generalization gap and the efficacy of the proposed method. 2. The scope of comparative studies is somewhat limited. The paper primarily compares with methods not designed for visual adaptation, which may not provide the most insightful comparison. It would have been advantageous to see how the proposed approach stacks up against other strategies specifically intended for visual adaptation. 3. An exploration of whether the proposed method can also benefit model-free algorithms such as Soft Actor-Critic (SAC) is missing. This could broaden the applicability of the findings and provide additional insights into the method's versatility. 4. It remains unclear why the proposed method underperforms with the Inverse Dynamics Model (IDM). A deeper exploration of this anomaly could bolster the robustness and reliability of the approach. 5. Finally, the similarity between the proposed method and the PAD method, which performs test-time domain adaptation in model-free RL, raises questions. The key differences are the use of a Spatial Transformer Network and the leveraging of the latent dynamics model of model-based methods instead of adding it as an auxiliary component. However, these differences, while noteworthy, do not necessarily represent a substantial departure from the PAD method. This similarity begs the question of whether the presented method offers significant novelty or if it is essentially an adjustment of existing approaches.

Questions

1. To better understand the importance of the generalization gap, could you provide performance data related to the original view? This information would help quantify the performance decrease when the agent is switched to a different view. 2. The use of inverse dynamics models with spatial transformer networks in your method isn't entirely clear. Does IDM completely replace the DM in the TD-MPC algorithm? Or does it just used as a side network during TD-MPC training and used for test adaptation only?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

2 fair

Limitations

The authors mentioned a limitation of this work: the proposed method has not been tested in real-world scenarios, such as with physical robots, which may limit its immediate applicability. Another potential limitation could be that the method has only been tested on model-based reinforcement learning (RL), whereas it would be relatively straightforward to evaluate its effectiveness in model-free RL.

Reviewer CdSm5/10 · confidence 4/52023-07-06

Summary

This paper focuses on improving the generalization ability of visual DRL to adapt to unseen views. The authors propose a model-based policy adaptation approach that combines spatial transformer networks with a self-supervised dynamics prediction objective to address this problem. The effectiveness of the approach is evaluated through experiments on three commonly used benchmarks.

Strengths

- This paper is well-written and easy to follow. - The proposed method is well-motivated. - The authors try to solve a realistic and essential problem, i.e., view generalization, especially in real-world or sim-to-real robotic scenarios.

Weaknesses

- While the problem setting is attractive, the technical contribution is insufficient as a full NeurIPS paper. It is a straightforward combination of multiple existing works, like STN. - There needs to be more than the current evaluation to demonstrate the superiority over existing works.

Questions

- The experiments need to be more comprehensive. The comparison baselines used are more like ablation studies rather than state-of-the-art (SOTA) methods, such as TD-MPC (the backbone of MoVie), DM (MoVie without STN), and IDM+STN (MoVie replacing DM with IDM). The authors should consider incorporating more commonly-used or SOTA data augmentation algorithms for visual DRL, such as DrQ-v2[1] and SEVA[2]. - The authors mention multiple times that only shallow STNs can improve performance, but they fail to provide deeper analysis, such as visual illustrations. Is the number of layers set for all tasks, or do different tasks require their own suitable layers? This information would be valuable. - The authors claim that "Our proposed method enables direct deployment of offline or simulation-trained agents ...", However, the lack of practical experimental demonstrations, such as robot manipulation, makes their claim unconvincing. I strongly suggest that the authors conduct real-world robot experiments to assess the applicability of their proposed method. References: - [1] Denis Yarats, Rob Fergus, Alessandro Lazaric, and Lerrel Pinto. Mastering visual continuous control: Improved data-augmented reinforcement learning. arXiv preprint arXiv:2107.09645, 2021. - [2] Nicklas Hansen, Hao Su, and Xiaolong Wang. Stabilizing deep q-learning with convnets and vision 362 transformers under data augmentation. Advances in Neural Information Processing Systems, 34, 2021.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Not Applicable.

Authorsrebuttal2023-08-19

Thank you for the review and awaiting your response

We sincerely thank you for your efforts in reviewing our paper and the suggestions again. We believe that we have resolved all the concerns mentioned in the review. Should there be any additional concerns, we are more than happy to address them! Thank you very much!

Reviewer CADg6/10 · confidence 4/52023-07-09

Summary

Adapting policy into new view setup is an important task in RL. This paper presents MoVie, Model-based policies for View generalization to achieve the fast view adaptation of model-based policy. Specifically, they combine spatial transformer networks into the encoder models and train it during test-time using dynamics prediction loss. Despite the simplicity, the experimental results show that the strong performance of MoVie over several benchmarks on various test set.

Strengths

1. The proposed method is simple yet provide stable performance gain over various methods. Experimental validation is extensive enough. 2. The paper is well written and easy to follow.

Weaknesses

1. Technical novelty and technical depth is limited. I might decrease scores if I found works that combine STN with test-time training. Besides, few discussions are made on the choice of test-time loss and entire architecture. For example, the experimental results suggest that the DM is better than IDM especially when combining test-time training, but the paper lacks discussion on why it is. Besides, I agree prediction loss is one of the natural loss to train STN, but I think it is also possible to extend the method to model-free methods by using some self-supervised loss function (like contrastive learning). 2. I think the experimental validation is already comprehensive, but it lacks qualitative analysis of their representation before/after adaptation. 3. No theoretical justification for the choice of test-time loss and or entire architecture.

Questions

See weakness section.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

N/A

Authorsrebuttal2023-08-19

Thank you for the review and awaiting your response

We sincerely thank you for your efforts in reviewing our paper and the suggestions again. We believe that we have resolved all the concerns mentioned in the review. Should there be any additional concerns, we are more than happy to address them! Thank you very much!

Reviewer N1Ux2023-08-11

I appreciate the authors' responses and clarifications regarding my questions. After considering their input, I still believe that this paper is of sufficient quality to be accepted for NeurIPS.

Authorsrebuttal2023-08-11

We thank the reviewer for acknowledging our efforts.

Reviewer CdSm2023-08-19

I appreciate the authors' additional experimental results and critical analysis. They addressed most of my concerns. So I decide to raise my final rating score. However, Regarding Q5, I acknowledge the importance of simulation but have concerns about your claim that "Our proposed method enables direct deployment of offline or simulation-trained agents" (Lines 186-189). Due to the significant difference (domain gap) in input images between the simulation and the real world, a sim-2-real module is usually needed to mitigate these discrepancies. I am skeptical about the feasibility of deploying your MoVie directly to real-world experiments. Thus, I recommend that the authors present concrete experimental findings instead of mere assertions.

Authorsrebuttal2023-08-20

Thank you for your constructive response and the acknowledgement of our effort. We agree that there is a difference in input images between the simulation and the real world; hence, real world experiment is beneficial. We would like to point out that our method can narrow the domain gap in input images to some extent. As shown in the table below, MoVie effectively facilitates the agent's adaptation to different visual domains. In this experiment, we introduce changes to the appearance of objects such as the ground, table, and background in the environment and report cumulative rewards for DMControl tasks and success rates for robotic manipulation tasks. While these being a surrogate setting, we also plan to conduct experiments in real world in our future work to further demonstrate that MoVie enables direct deployment of offline or simulation-trained agents. | Tasks | TD-MPC (original) | TD-MPC under appearance change | MoVie under appearance change | | :-----: | :-----: | :-----: | :-----: | | Walker, walk | $ 944.99\pm21.71 $ | $ 589.76\pm53.27 $ | $ 882.00\pm72.68 $ | | xArm, reach | $ 0.96\pm0.05 $ | $ 0.15\pm0.05 $ | $ 0.78\pm0.02 $ |

Reviewer CADg2023-08-20

I would like to keep my rating

Thank you for your detailed response. I still think the paper provides a good contribution to the community, and therefore keep my rating.

Reviewer E44Y2023-08-21

I appreciate the authors' response. They addressed most of my concerns. I decide to raise my rating score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC