Pith. sign in

REVIEW 1 major objections 1 minor 55 references

Task-Relevant Representation Decoupling for Visual Reinforcement Learning Generalization

T0 review · 1 major / 1 minor · reviewed 2026-07-02 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A self-supervised method decouples visual observations into task-relevant and irrelevant parts to improve reinforcement learning generalization.

desk verdict T2RD adds cross-dynamic prediction on top of content-style decoupling to target task-relevant features in visual RL, but the abstract leaves the isolation claim without direct support. read the letter →

arxiv 2607.00796 v1 pith:UXIPHJ3X submitted 2026-07-01 cs.LG

Jinwen Wang , Youfang Lin , Xiaobo Hu , Qian Xu , Shuo Wang , Zhuo Chen , Kai Lv This is my paper
classification cs.LG
keywords visualreinforcementlearningrepresentationdecouplinggeneralizationself-supervisedtask-relevantfeaturesDeepMindControlSuite
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes separating visual inputs in reinforcement learning into features that matter for the task and those that do not. This separation relies on three components: enforcing consistency in task-relevant representations, cross-reconstruction between content and style, and cross-dynamic prediction to refine relevance. The resulting algorithm focuses learning on the useful parts of the observation, which allows policies to transfer to environments with different backgrounds or visual styles. If successful, this reduces overfitting to training-specific details that do not affect the underlying control problem.

What carries the argument

Task-Relevant Representation Decoupling (T2RD) algorithm, which applies three self-supervised components to isolate task-relevant features from visual observations.

What would settle it

A controlled test where T2RD fails to improve generalization metrics over baseline visual RL methods on held-out environments in the DeepMind Control Suite.

Watch

Extended reading notes

Core claim

The T2RD algorithm decouples observations into task-relevant and task-irrelevant representations through task-relevant representation consistency, cross-reconstruction, and cross-dynamic prediction, with the third component ensuring content features are task-relevant rather than merely style-invariant.

Load-bearing premise

The cross-dynamic prediction component can reliably isolate task-relevant features from the content representations produced by the first two components.

Editorial extensions

If this is right

  • Policies trained this way transfer to environments with altered visual styles without retraining.
  • Sample efficiency rises because the agent ignores features unrelated to the control objective.
  • The same decoupling applies across both simulation control tasks and robotic manipulation settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Similar separation of relevant and irrelevant visual signals could reduce reliance on explicit domain randomization in other visual control problems.
  • The approach might extend to settings where dynamics change over time by re-applying the dynamic prediction step at deployment.
  • If the components prove stable, the method could serve as a modular add-on to existing visual encoders without changing the downstream policy architecture.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 1 minor

Summary. The paper proposes Task-Relevant Representation Decoupling (T2RD), a self-supervised algorithm with three components—task-relevant representation consistency, cross-reconstruction, and cross-dynamic prediction—to decouple observations into task-relevant and task-irrelevant representations for improved generalization in visual reinforcement learning. It claims state-of-the-art generalization performance and sample efficiency on the DeepMind Control Suite and Robotic Manipulation tasks.

Significance. If the cross-dynamic prediction step reliably isolates task-relevant dynamics as claimed, the method would offer a concrete mechanism for mitigating overfitting to task-irrelevant visual features, potentially advancing generalization techniques in visual RL beyond current contrastive or reconstruction-based approaches.

major comments (1)
  1. [Method section on the three components] The description of the third component (cross-dynamic prediction) states that the first two components produce content representations that 'are not necessarily task-relevant' and that component 3 is introduced 'to further refine task-relevant features from content representations.' No ablation study isolates the incremental effect of component 3 on task-relevance (as opposed to regularization or auxiliary loss effects), nor is any diagnostic provided that compares task-relevance of content features before versus after this component (e.g., via linear probing on task labels or controlled environment variants). This assumption is load-bearing for the decoupling premise and the SOTA generalization claim.
minor comments (1)
  1. [Abstract] The abstract asserts SOTA results and sample efficiency without reporting any quantitative metrics, baselines, number of runs, or statistical significance; even a concise abstract should include at least the key performance deltas.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive comments on the method description. We address the major comment below and agree that additional analyses are warranted to strengthen the claims.

read point-by-point responses

  1. Referee: [Method section on the three components] The description of the third component (cross-dynamic prediction) states that the first two components produce content representations that 'are not necessarily task-relevant' and that component 3 is introduced 'to further refine task-relevant features from content representations.' No ablation study isolates the incremental effect of component 3 on task-relevance (as opposed to regularization or auxiliary loss effects), nor is any diagnostic provided that compares task-relevance of content features before versus after this component (e.g., via linear probing on task labels or controlled environment variants). This assumption is load-bearing for the decoupling premise and the SOTA generalization claim.

    Authors: We agree that the current manuscript lacks an ablation that isolates the incremental contribution of cross-dynamic prediction specifically to task-relevance (distinct from general regularization or auxiliary loss benefits), as well as direct diagnostics such as linear probing on task labels or controlled environment variants to compare content features before versus after component 3. While the end-to-end generalization results support the overall approach, these targeted studies would provide clearer validation of the load-bearing assumption. In the revised manuscript we will add (i) an ablation removing only component 3 and (ii) linear probing diagnostics on task-relevant labels across environment variants to quantify the refinement effect. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: method is a proposed self-supervised algorithm without self-referential derivations

full rationale

The paper introduces T2RD as a three-component self-supervised algorithm for representation decoupling in VRL. The abstract and description outline task-relevant consistency, cross-reconstruction, and cross-dynamic prediction without any equations, fitted parameters renamed as predictions, or load-bearing self-citations. The decoupling premise and refinement step are presented as design choices, with SOTA claims resting on empirical results in DMC and manipulation tasks rather than any derivation that reduces to its own inputs by construction. No self-definitional, fitted-input, or ansatz-smuggling patterns appear.

Assumptions & free parameters 0 free parameters · 0 assumptions · 0 invented entities

Abstract-only review yields no visible free parameters, axioms, or invented entities; all such elements would require the methods and equations sections.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Task-Relevant Representation Decoupling for Visual Reinforcement Learning Generalization." pith.science (2026). https://pith.science/paper/UXIPHJ3X

@misc{pith2026260700796,
  author       = {Pith},
  title        = {Pith review of: Task-Relevant Representation Decoupling for Visual Reinforcement Learning Generalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UXIPHJ3X}},
  note         = {Machine review of arXiv:2607.00796}
}
read the original abstract

Visual Reinforcement Learning (VRL) has achieved considerable success in solving control tasks. However, generalizing learned policies to new environments remains a major challenge, as agents often overfit to task-irrelevant features in the training environment. To solve this problem, we introduce the concept of decoupling observations into task-relevant and task-irrelevant representations. Building on this idea, we propose a self-supervised Task-Relevant Representation Decoupling (T2RD) algorithm for VRL. This algorithm consists of three components: task-relevant representation consistency, cross-reconstruction, and cross-dynamic prediction. The first two components achieve the decoupling of content and style features, but the resulting content representations are not necessarily task-relevant. To further refine task-relevant features from content representations, we design the third component that introduces dynamic prediction. T2RD achieves State-Of-The-Art (SOTA) generalization performance and sample efficiency in the DeepMind Control Suite and Robotic Manipulation tasks.

Figures

Figures reproduced from arXiv: 2607.00796 by the authors.

Figure 1
Figure 1. The core idea of our method. We decouple the visual observation into task-relevant representation and task-irrelevant representation. Task-relevant representations have domain-invariant properties, allowing them to generalize well across different domains. utilized to support learning policies. This decoupling idea has two advantages: 1) The representations supporting policy learning are expected to exclude task-irr… view at source ↗
Figure 2
Figure 2. Overview of our model. Our model includes two branches: the online branch and the target branch. T2RD consists of three components: task-relevant representation consistency, cross-reconstruction, and cross￾dynamic prediction. Our objective is to achieve the decoupling of task-relevant and task-irrelevant features through the above three modules. Subsequently, only the task-relevant features are used for policy learn… view at source ↗
Figure 3
Figure 3. Policy Learner. The observation 𝑜𝑡 is decoupled into the task-relevant representation 𝜇𝑡 and the task-irrelevant representation 𝜖𝑡 through an encoder 𝑓𝜃 and a dual-head projection 𝑔𝜃 . Then, only the task￾relevant representation 𝜇𝑡 is utilized for policy learning. We also implement a cross-domain prediction technique here. Specifically, we use the task￾relevant representation (𝜇 ′ 𝑡 ) from the augmented view and the… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Examples of data augmentation. We separately present the original observation and the results after applying two data augmentation techniques: Random Overlay and Random Convolution. Baselines. We compare our approach with other SOTA algorithms: 1) DrQ [44] applies data…
Figure 5
Figure 5. Figure 5: Examples of DMControl-GB. We separately present the training example observations and three test settings: “Color Hard”, “Video Easy” and “Video Hard” [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Sample efficiency. We visualize the training curves of T2RD, SGQN, BiT, and TRP. The solid line represents the mean reported, and the shading represents the standard deviation. It can be seen that T2RD has the best sample efficiency and asymptotic performance. T2RD has…
Figure 7
Figure 7. Figure 7: Test curves on DMControl-GB. We compare the test curves of T2RD with SGQN, BiT, and TRP on three tasks in DMControl-GB: “Cartpole Swingup”, “Finger Spin”, and “Walker Stand”, across three different environmental settings: Color Hard, Video Easy, and Video Hard. T2RD (O…
Figure 8
Figure 8. Figure 8: t-SNE of representation. We visualize the representations learned from T2RD, SGQN, PIEG, SVEA. T2RD exhibits the best intra-class cohesion and inter-class separation. T2RD successfully learns task-relevant representations. In [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: t-SNE visualization of decoupled task-relevant and task-irrelevant features across training steps [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Cross-reconstruction visualization. We use the task-relevant features (content) 𝜇 ′ 𝑡 from the augmented observation 𝑜 ′ 𝑡 and the task-irrelevant features (style) 𝜖𝑡 from the original observation to perform cross-reconstruction, obtaining the reconstructed image 𝑜ˆ𝑡 …
Figure 11
Figure 11. Figure 11: Visualizations of attention maps. We visualize the attention areas of our algorithm under different backgrounds in DMControl-GB. Easy”, and “Video Hard”. T2RD can accurately perceive task-relevant regions in visual observations, regardless of the presence of backgroun…
Figure 12
Figure 12. Figure 12: Examples of Robotic Manipulation. We present training example observations and four test settings for the “Reach” and “Peg in box” tasks [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: Ablation Study on DMControl-GB. We conduct ablation studies on the “Cartpole Swingup” and “Finger Spin” tasks under the “Video Hard” setting. “T2RD (w/o reconstruction)” indicates the exclusion of the cross-reconstruction module; “T2RD (w/o dynamic prediction)” denote…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 55 canonical work pages

  1. [1]

    David Bertoin, Adil Zouitine, Mehdi Zouitine, and Emmanuel Rachelson. 2022. Look where you look! Saliency-guided Q-networks for generalization in visual Reinforcement Learning. Advances in Neural Information Processing Systems 35 (2022), 30693–30706

    work page 2022

  2. [2]

    Yevgen Chebotar, Quan Vuong, Karol Hausman, Fei Xia, Yao Lu, Alex Irpan, Aviral Kumar, Tianhe Yu, Alexander Herzog, Karl Pertsch, et al. 2023. Q-transformer: Scalable offline reinforcement learning via autoregressive q-functions. In Conference on Robot Learning . PMLR, 3909–3928

    work page 2023

  3. [3]

    Jianda Chen and Sinno Jialin Pan. 2022. Learning Representations via a Robust Behavioral Metric for Deep Reinforce- ment Learning. In Advances in Neural Information Processing Systems 35 (2022)

    work page 2022

  4. [4]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. ImageNet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2009 . IEEE, 248–255

    work page 2009

  5. [5]

    Hehe Fan, Yi Yang, and Mohan Kankanhalli. 2022. Point spatio-temporal transformer networks for point cloud video modeling. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 2 (2022), 2181–2192

    work page 2022

  6. [6]

    Hehe Fan, Liang Zheng, Chenggang Yan, and Yi Yang. 2018. Unsupervised person re-identification: Clustering and fine-tuning. ACM Transactions on Multimedia Computing, Communications, and Applications 14, 4 (2018), 1–18

    work page 2018

  7. [7]

    Linxi Fan, Guanzhi Wang, De-An Huang, Zhiding Yu, Li Fei-Fei, Yuke Zhu, and Animashree Anandkumar. 2021. SECANT: Self-Expert Cloning for Zero-Shot Generalization of Visual Policies. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021 (Proceedings of Machine Learning Research, Vol. 139) . PMLR, 3088–3099

    work page 2021

  8. [8]

    Norm Ferns, Prakash Panangaden, and Doina Precup. 2011. Bisimulation Metrics for Continuous Markov Decision Processes. SIAM J. Comput. 40, 6 (2011), 1662–1714

    work page 2011

Show all 55 references
  1. [9]

    Abel Gonzalez-Garcia, Joost van de Weijer, and Yoshua Bengio. 2018. Image-to-image translation for cross-domain disentanglement. In Advances in Neural Information Processing Systems 31 (2018) . 1294–1305

    2018

  2. [10]

    Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko

    Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. 2020. Bootstrap Your Own Latent -...

    2020

  3. [11]

    Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. 2018. Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor. In Proceedings of the 35th International Conference on Machine Learning, ICML 2018 (Proceedings of Machine L...

    2018

  4. [12]

    Efros, Lerrel Pinto, and Xiaolong Wang

    Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A. Efros, Lerrel Pinto, and Xiaolong Wang. 2021. Self-Supervised Policy Adaptation during Deployment. In 9th International Conference on Learning Representations, ICLR 2021. OpenReview.net

    2021

  5. [13]

    Nicklas Hansen, Hao Su, and Xiaolong Wang. 2021. Stabilizing Deep Q-Learning with ConvNets and Vision Trans- formers under Data Augmentation. In Advances in Neural Information Processing Systems 34 (2021) . 3680–3693

    2021

  6. [14]

    Nicklas Hansen and Xiaolong Wang. 2021. Generalization in Reinforcement Learning by Soft Data Augmentation. In 2021 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 13611–13617

    2021

  7. [15]

    Xiaobo Hu, Youfang Lin, Yue Liu, Jinwen Wang, Shuo Wang, Hehe Fan, and Kai Lv. 2025. A Reliable Representation with Bidirectional Transition Model for Visual Reinforcement Learning Generalization. ACM Transactions on Multimedia Computing, Communications and Applications (2025)

    2025

  8. [16]

    Xiaobo Hu, Youfang Lin, Jinwen Wang, Yue Liu, Shuo Wang, Hehe Fan, and Kai Lv. 2025. Bidirectional transition consistency between multi-domain observations for visual reinforcement learning generalization. Neural Networks (2025), 108265

    2025

  9. [17]

    Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. 2023. VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models. arXiv preprint arXiv:2307.05973 (2023)

    2023 arXiv

  10. [18]

    Rishabh Jangir, Nicklas Hansen, Sambaran Ghosal, Mohit Jain, and Xiaolong Wang. 2022. Look Closer: Bridging Egocentric and Third-Person Views With Transformers for Robotic Manipulation. IEEE Robotics Autom. Lett. 7, 2 (2022), 3046–3053. 22 Jinwen Wang et al

    2022

  11. [19]

    Michael Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. 2020. Reinforcement Learning with Augmented Data. In Advances in Neural Information Processing Systems 33 (2020)

    2020

  12. [20]

    Michael Laskin, Aravind Srinivas, and Pieter Abbeel. 2020. CURL: Contrastive Unsupervised Representations for Reinforcement Learning. InProceedings of the 37th International Conference on Machine Learning, ICML 2020 (Proceedings of Machine Learning Research, Vol. 119) . PMLR, ...

    2020

  13. [21]

    Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. 2016. End-to-End Training of Deep Visuomotor Policies. J. Mach. Learn. Res. 17 (2016), 39:1–39:40

    2016

  14. [22]

    Lu Li, Jiafei Lyu, Guozheng Ma, Zilin Wang, Zhenjie Yang, Xiu Li, and Zhiheng Li. 2023. Normalization Enhances Generalization in Visual Reinforcement Learning. arXiv preprint arXiv:2306.00656 (2023)

    2023

  15. [23]

    Anthony Liang, Jesse Thomason, and Erdem Biyik. 2024. ViSaRL: Visual Reinforcement Learning Guided by Human Saliency. In IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2024, . IEEE, 2907–2912

    2024

  16. [24]

    Dominik Lorenz, Leonard Bereska, Timo Milbich, and Björn Ommer. 2019. Unsupervised Part-Based Disentangling of Object Shape and Appearance. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019 . Computer Vision Foundation / IEEE, 10955–10964

    2019

  17. [25]

    Wenhan Luo, Peng Sun, Fangwei Zhong, Wei Liu, Tong Zhang, and Yizhou Wang. 2020. End-to-End Active Object Tracking and Its Real-World Deployment via Reinforcement Learning. IEEE Trans. Pattern Anal. Mach. Intell. 42, 6 (2020), 1317–1332

    2020

  18. [26]

    Michaël Mathieu, Junbo Jake Zhao, Pablo Sprechmann, Aditya Ramesh, and Yann LeCun. 2016. Disentangling factors of variation in deep representation using adversarial training. In Advances in Neural Information Processing Systems 29 (2016). 5041–5049

    2016

  19. [27]

    Rusu, Joel Veness, Marc G

    Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Mar- tin A. Riedmiller, Andreas Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra...

    2015

  20. [28]

    Masashi Okada and Tadahiro Taniguchi. 2021. Dreaming: Model-based Reinforcement Learning by Latent Imagination without Reconstruction. In IEEE International Conference on Robotics and Automation, ICRA 2021 . IEEE, 4209–4215

    2021

  21. [29]

    Xuanchi Ren, Tao Yang, Yuwang Wang, and Wenjun Zeng. 2021. Rethinking Content and Style: Exploring Bias for Unsupervised Disentanglement. In IEEE/CVF International Conference on Computer Vision Workshops, ICCVW 2021 . IEEE, 1823–1832

    2021

  22. [30]

    Eduardo Hugo Sanchez, Mathieu Serrurier, and Mathias Ortner. 2020. Learning Disentangled Representations via Mutual Information Estimation. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII (Lecture Notes in Comp...

    2020

  23. [31]

    Akanksha Saran, Ruohan Zhang, Elaine Schaertl Short, and Scott Niekum. 2021. Efficiently Guiding Imitation Learning Agents with Human Gaze. In AAMAS ’21: 20th International Conference on Autonomous Agents and Multiagent Systems, Virtual Event, United Kingdom, May 3-7, 2021 . A...

    2021

  24. [32]

    Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdolmaleki, Josh Merel, Andrew Lefrancq, et al. 2018. Deepmind control suite. arXiv preprint arXiv:1801.00690 (2018)

    2018 arXiv

  25. [33]

    Tenenbaum and William T

    Joshua B. Tenenbaum and William T. Freeman. 2000. Separating Style and Content with Bilinear Models. Neural Comput. 12, 6 (2000), 1247–1283

    2000

  26. [34]

    Chu-ran Wang, Fei Gao, Fandong Zhang, Fangwei Zhong, Yizhou Yu, and Yizhou Wang. 2022. Disentangling Disease- related Representation from Obscure for Disease Prediction. InInternational Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA (Proce...

    2022

  27. [35]

    Jinwen Wang, Youfang Lin, Xiaobo Hu, Siyu Yang, Sheng Han, Shuo Wang, and Kai Lv. 2025. From Pixels to Temporal Correlations: Learning Informative Representations for Reinforcement Learning Pre-training. In Proceedings of the 33rd ACM International Conference on Multimedia . 679–688

    2025

  28. [36]

    Shuo Wang, Zhihao Wu, Xiaobo Hu, Youfang Lin, and Kai Lv. 2023. Skill-Based Hierarchical Reinforcement Learning for Target Visual Navigation. IEEE Transactions on Multimedia 25 (2023), 8920–8932

    2023

  29. [37]

    Shuo Wang, Zhihao Wu, Xiaobo Hu, Jinwen Wang, Youfang Lin, and Kai Lv. 2024. What Effects the Generalization in Visual Reinforcement Learning: Policy Consistency with Truncated Return Prediction. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024 . AAAI Pre...

    2024

  30. [38]

    Shuo Wang, Zhihao Wu, Jinwen Wang, Xiaobo Hu, Youfang Lin, and Kai Lv. 2024. How to Learn Domain-Invariant Representations for Visual Reinforcement Learning: An Information-Theoretical Perspective. In Proceedings of the Thirty-Third International Joint Conference on Artificial...

    2024

  31. [39]

    Xudong Wang, Long Lian, and Stella X. Yu. 2021. Unsupervised Visual Attention and Invariance for Reinforcement Learning. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021 . 6677–6687

    2021

  32. [40]

    Ziyu Wang, Yanjie Ze, Yifei Sun, Zhecheng Yuan, and Huazhe Xu. 2023. Generalizable Visual Reinforcement Learning with Segment Anything Model. arXiv preprint arXiv:2312.17116 (2023). Task-Relevant Representation Decoupling for Visual Reinforcement Learning Generalization 23

    2023

  33. [41]

    Wayne Wu, Kaidi Cao, Cheng Li, Chen Qian, and Chen Change Loy. 2019. Disentangling Content and Style via Unsupervised Geometry Distillation. In Deep Generative Models for Highly Structured Data, ICLR 2019 Workshop . OpenReview.net

    2019

  34. [42]

    Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. 2021. Robust and Generalizable Visual Representation Learning via Random Convolutions. In 9th International Conference on Learning Representations, ICLR

    2021

  35. [43]

    Rui Yang, Jie Wang, Qijie Peng, Ruibo Guo, Guoping Wu, and Bin Li. 2025. Learning Robust Representations with Long-Term Information for Generalization in Visual Reinforcement Learning. InThe Thirteenth International Conference on Learning Representations, ICLR 2025 . OpenReview.net

    2025

  36. [44]

    Denis Yarats, Ilya Kostrikov, and Rob Fergus. 2021. Image Augmentation Is All You Need: Regularizing Deep Reinforce- ment Learning from Pixels. In 9th International Conference on Learning Representations, ICLR 2021 . OpenReview.net

    2021

  37. [45]

    Dacheng Yin, Xuanchi Ren, Chong Luo, Yuwang Wang, Zhiwei Xiong, and Wenjun Zeng. 2022. Retriever: Learning Content-Style Representation as a Token-Level Bipartite Graph. In The Tenth International Conference on Learning Representations, ICLR 2022. OpenReview.net

    2022

  38. [46]

    Zhecheng Yuan, Guozheng Ma, Yao Mu, Bo Xia, Bo Yuan, Xueqian Wang, Ping Luo, and Huazhe Xu. 2022. Don’t Touch What Matters: Task-Aware Lipschitz Data Augmentation for Visual Reinforcement Learning. InProceedings of the Thirty-First International Joint Conference on Artificial ...

    2022

  39. [47]

    Zhecheng Yuan, Zhengrong Xue, Bo Yuan, Xueqian Wang, Yi Wu, Yang Gao, and Huazhe Xu. 2022. Pre-Trained Image Encoder for Generalizable Visual Reinforcement Learning. In Advances in Neural Information Processing Systems 35 (2022)

    2022

  40. [48]

    Amy Zhang, Rowan Thomas McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine. 2021. Learning Invariant Representations for Reinforcement Learning without Reconstruction. In 9th International Conference on Learning Representations, ICLR 2021. OpenReview.net

    2021

  41. [49]

    Ruihan Zhang and Jun Sun. 2024. Certified robust accuracy of neural networks are bounded due to Bayes errors. In International Conference on Computer Aided Verification . Springer, 352–376

    2024

  42. [50]

    Ruijie Zheng, Xiyao Wang, Yanchao Sun, Shuang Ma, Jieyu Zhao, Huazhe Xu, Hal Daumé III, and Furong Huang. 2023. TACO: Temporal Latent Action-Driven Contrastive Loss for Visual Reinforcement Learning. InAdvances in Neural Information Processing Systems 36 (2023)

    2023

  43. [51]

    Fangwei Zhong, Peng Sun, Wenhan Luo, Tingyun Yan, and Yizhou Wang. 2021. Towards Distraction-Robust Active Visual Tracking. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021 (Proceedings of Machine Learning Research, Vol. 139), Marina Meila and...

    2021

  44. [52]

    Fangwei Zhong, Kui Wu, Hai Ci, Churan Wang, and Hao Chen. 2024. Empowering Embodied Visual Tracking with Visual Foundation Models and Offline RL. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LXXIII (Le...

    2024

  45. [53]

    Bolei Zhou, Àgata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. 2018. Places: A 10 Million Image Database for Scene Recognition. IEEE Trans. Pattern Anal. Mach. Intell. 40, 6 (2018), 1452–1464

    2018

  46. [54]

    Bohan Zhou, Ke Li, Jiechuan Jiang, and Zongqing Lu. 2023. Learning from Visual Observation via Offline Pretrained State-to-Go Transformer. arXiv preprint arXiv:2306.12860 (2023)

    2023

  47. [55]

    Yiming Zuo, Weichao Qiu, Lingxi Xie, Fangwei Zhong, Yizhou Wang, and Alan L. Yuille. 2019. CRAVES: Controlling Robotic Arm With a Vision-Based Economic System. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019. Computer Vision Foundation / IEEE, 4214–4223

    2019

Pith tools

Reviewed July 2, 2026 · model on record in the stance chip above.