pith. sign in

arxiv: 2607.01897 · v1 · pith:MRGE5JZRnew · submitted 2026-07-02 · 💻 cs.LG · cs.AI

Rank-Then-Act: Reward-Free Control from Frame-Order Progress

Yuriy Maksyuta , George Bredis , Ruslan Rakhimov , Daniil Gavrilov This is my paper

Pith reviewed 2026-07-03 17:40 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords reward-free reinforcement learningvision-language modelsordinal progress scoringrank correlation rewardsvideo demonstrationsgroup relative policy optimizationpolicy learning from videoscross-task transfer
0
0 comments X

The pith

A vision-language model trained on shuffled video frames supplies rank-correlation rewards sufficient to learn control policies without environment rewards.

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

The paper introduces Rank-Then-Act, which learns policies from expert videos alone by first training a vision-language model offline to score ordinal progress. The training uses a Group Relative Policy Optimization objective over shuffled frame sequences so the model must recover order from visual semantics. Instead of treating the model's scores as direct rewards, the method computes Spearman rank correlation between the predicted progress rankings and actual temporal indices inside each interaction window, producing a bounded and scale-invariant signal. This signal then guides reinforcement learning, and the same pretrained scorer transfers across tasks.

Core claim

Training a vision-language model with Group Relative Policy Optimization over shuffled frame sequences creates an ordinal progress scorer; feeding the Spearman rank correlation between its predictions and true frame times into reinforcement learning then yields effective policies from video demonstrations without environment rewards, matching or exceeding prior video-based methods on PyBoy, PointMaze, and MetaWorld while enabling cross-task reuse of one scorer.

What carries the argument

The Spearman rank correlation between the VLM's predicted progress rankings and the true temporal indices of frames inside each interaction window, used as the reward signal.

If this is right

  • Policies can be trained without access to or design of any environment reward function.
  • One progress scorer pretrained on videos transfers to new tasks and environments without retraining.
  • The correlation reward remains bounded and scale-invariant, avoiding calibration issues common in direct score-based rewards.
  • The same framework applies to both discrete control benchmarks and continuous control tasks with comparable or better results than earlier video reward methods.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar ordinal video supervision could extend to sequential problems outside robotic control where reward specification is costly.
  • Testing the progress scorer on videos recorded under varied lighting or camera motion would reveal how much the method depends on clean visual semantics.
  • The approach might combine with other forms of weak supervision such as language instructions to further reduce demonstration requirements.

Load-bearing premise

The Group Relative Policy Optimization objective on shuffled sequences forces the vision-language model to recover temporal order from visual semantics rather than from any trivial time cues in the frames.

What would settle it

A controlled test in which the VLM achieves high accuracy on ordering shuffled held-out sequences yet the resulting correlation-based rewards produce no policy improvement during reinforcement learning, or in which replacing visual frames with time-stamped but visually identical frames still yields high GRPO performance.

Figures

Figures reproduced from arXiv: 2607.01897 by Daniil Gavrilov, George Bredis, Ruslan Rakhimov, Yuriy Maksyuta.

Figure 1
Figure 1. Figure 1: RTA: reward-free control in two stages. Stage 1 trains a VLM progress scorer on shuffled clips with a GRPO objective that maximizes progress–time Spearman ρ. Stage 2 uses the frozen scorer online: every N steps, we score a window (with L shuffles) and use the average ρ ∈ [−1, 1] as the only reward for PPO-style learning. 3.2 Stage 1: Listwise Progress Scorer via GRPO Expert clip batching. Given an expert v… view at source ↗
Figure 2
Figure 2. Figure 2: Cross-level transfer of or￾dinal progress. Heatmap of validation progress–time ρ for source (rows) → tar￾get (columns). Diagonals are in-level; off￾diagonals show transfer. Pooled training (all levels or diverse games) transfers best; sim￾pler Level 1 generalizes from most sources. VLM-based scorer trained on a single level (or YouTube dataset) recovers progress from shuffled sequences across different lev… view at source ↗
Figure 3
Figure 3. Figure 3: PointMaze-UMaze results. Results in this table report mean ± std across 5 seeds, smoothed with a moving average (window size 5) for visual clarity. RTA outperforms R2R both with and without GAIL. The oracle reward is included for comparison, although it is not directly comparable to our method. 0.0 0.2 0.4 0.6 0.8 1.0 Step 0 100 200 300 400 500 episode return 1e6 door-open 0.0 0.2 0.4 0.6 0.8 1.0 Step 100 … view at source ↗
Figure 4
Figure 4. Figure 4: MetaWorld different tasks results. Stage-1 RTA is trained once on a mixed dataset of demonstrations, while R2R is trained separately per task. We additionally evaluate Stage-1 scorers trained on different data sources, including COIN and YouTube datasets. Results report mean ± std over 4 seeds. RTA + GAIL achieves comparable returns to R2R across all tasks, and outperforms it on door-open and door-close. U… view at source ↗
Figure 5
Figure 5. Figure 5: Effect of window size on the reward signal on cyclic trajectory. Small windows produce increasing reward despite no true progress, reflecting local consistency. Medium windows better capture the absence of global progress, while large windows yield near-zero signal due to averaging. All window sizes assign high reward to the final segment where the goal is achieved. 5 Discussion By learning policies direct… view at source ↗
Figure 6
Figure 6. Figure 6: LOOP benefits from starting-point refreshing. Without refreshing, rewards improve but may not reach success when N is shorter than the task. Refreshing from the best terminal state after threshold τ enables completion; brief post-refresh dips arise if N exceeds remaining steps. 25 50 75 100 125 150 175 200 Training step 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 Reward lvl-1 lvl-2 lvl-3 lvl-4 lvl-5 lvl-6… view at source ↗
Figure 7
Figure 7. Figure 7: Stage 1 learns reliable ordinal progress. Left: per-level Catrap training shows fast rises in progress–time ρ with slower convergence on intricate levels. Right: training on full playthroughs across games also converges, with slightly lower plateaus due to non-informative segments (e.g., menus). A Appendix A.1 Experiments with LOOP For experiments with LOOP-an extension of GRPO to multi-turn settings-we se… view at source ↗
Figure 8
Figure 8. Figure 8: Example of good level completion with well-trained stage 1 model with reward 0.77. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Example of bad level compeletion with well-trained stage 1 model with reward -0.04. [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Example of bad level compeletion with undertrained stage 1 model with reward 0.66 [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Example of good level compeletion with undertrained stage 1 model with reward 0.15 [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Correlation-only reward drives control. For Catrap L2 mean accumulated VOC-reward during training highly correlates with the success rate over episode [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Level 3 exposes training sensitivity. Over 400 GRPO steps, progress–time ρ alternates between gains and dips, indicating that ambiguous content learns but benefits from careful data prep and sufficient duration. 0 20 40 60 80 Total time, sec 0 100 200 300 400 500 Episode return 1e3 door-open 0 20 40 60 Total time, sec 100 200 300 400 500 600 Episode return 1e3 drawer-open 0 20 40 60 80 Total time, sec 0 1… view at source ↗
Figure 14
Figure 14. Figure 14: Training efficiency of RTA compared to Rank2Reward on MetaWorld tasks under aligned [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Training efficiency of RTA compared to Rank2Reward on UMaze under aligned computa [PITH_FULL_IMAGE:figures/full_fig_p020_15.png] view at source ↗
read the original abstract

We introduce Rank-Then-Act (RTA), a framework for learning control policies from expert video demonstrations without environment rewards. RTA trains a Vision-Language Model (VLM) offline as a progress-based ordinal scorer, using a Group Relative Policy Optimization (GRPO) objective over shuffled frame sequences, which forces the model to recover temporal ordering from visual semantics rather than trivial time cues. Importantly, instead of using the scorer directly as a scalar reward model, we propose a correlation-based reward function for reinforcement learning: at each interaction window, we compute the Spearman rank correlation between predicted progress rankings and true temporal indices, yielding a bounded, scale-invariant learning signal. This design decouples reward learning from absolute calibration and enables stable transfer across tasks and environments. We evaluate RTA on discrete control benchmarks (PyBoy: Catrap, Kirby) and continuous control tasks (PointMaze, MetaWorld). RTA consistently matches or outperforms prior video-based reward learning methods and rank-based baselines, while demonstrating strong cross-task reuse of a single pretrained progress scorer. Our results suggest that correlation-structured supervision over video-derived ordinal signals is sufficient for policy learning, offering a scalable alternative to explicit reward design.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper introduces Rank-Then-Act (RTA), a reward-free framework for learning control policies from expert video demonstrations. A VLM is trained offline via Group Relative Policy Optimization (GRPO) on shuffled frame sequences to produce a progress-based ordinal scorer. This scorer is not used directly as a reward; instead, at each interaction window the Spearman rank correlation between the predicted progress rankings and the true temporal indices supplies a bounded, scale-invariant reward signal for RL. The method is evaluated on discrete control benchmarks (PyBoy: Catrap, Kirby) and continuous control tasks (PointMaze, MetaWorld), with claims of consistent outperformance over prior video-based reward learning methods and rank-based baselines, plus strong cross-task reuse of a single pretrained scorer.

Significance. If the central claim holds, the work demonstrates that correlation-structured supervision derived from video-derived ordinal signals can suffice for policy learning, providing a scalable alternative to explicit reward design. The decoupling of reward learning from absolute calibration via Spearman correlation is a potentially useful design choice for transfer.

major comments (2)
  1. [Abstract] Abstract: the claim that GRPO over shuffled frame sequences 'forces the model to recover temporal ordering from visual semantics rather than trivial time cues' is load-bearing for the transferability of the correlation reward, yet no mechanism is described that would provably eliminate consistent low-level signals (e.g., frame compression patterns, lighting gradients, or motion blur statistics) that survive shuffling. If such cues remain, high GRPO scores can be achieved without learning transferable ordinal semantics.
  2. [Abstract] Abstract: the assertion of 'consistent outperformance' on the listed benchmarks supplies no quantitative results, error bars, ablation details, or verification of the central claim, rendering the empirical support unassessable from the provided text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading and constructive feedback. We respond to each major comment below and indicate where revisions will be made.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that GRPO over shuffled frame sequences 'forces the model to recover temporal ordering from visual semantics rather than trivial time cues' is load-bearing for the transferability of the correlation reward, yet no mechanism is described that would provably eliminate consistent low-level signals (e.g., frame compression patterns, lighting gradients, or motion blur statistics) that survive shuffling. If such cues remain, high GRPO scores can be achieved without learning transferable ordinal semantics.

    Authors: We acknowledge the validity of this observation. The abstract phrasing asserts a stronger guarantee than the GRPO objective and shuffling procedure can formally deliver, as low-level persistent cues are not provably ruled out. We will revise the abstract to use more cautious language (e.g., 'encourages the model to recover temporal ordering primarily from visual semantics') and will add a short paragraph in the discussion or limitations section noting this possibility while pointing to the observed cross-task transfer as empirical evidence of practical utility. No additional experiments are feasible within the revision timeline. revision: partial

  2. Referee: [Abstract] Abstract: the assertion of 'consistent outperformance' on the listed benchmarks supplies no quantitative results, error bars, ablation details, or verification of the central claim, rendering the empirical support unassessable from the provided text.

    Authors: The abstract is intentionally concise; all quantitative results (means, standard deviations across seeds, and ablations) appear in the experimental section, tables, and figures of the full manuscript. To improve standalone readability we will revise the abstract to include one or two key quantitative highlights (e.g., average improvement margins) subject to length limits. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation uses external ground-truth indices and offline GRPO training on separate expert videos

full rationale

The paper trains a VLM progress scorer offline via GRPO on shuffled expert video frames, then defines the RL reward as Spearman correlation between the scorer's output ranks and the true temporal indices of the agent's current interaction window. The true indices are external to the model and supplied by the environment interaction log; the scorer parameters are fitted on a disjoint set of demonstration sequences. No equation reduces a claimed prediction to its own fitted values by construction, no self-citation chain is invoked to justify uniqueness or an ansatz, and the correlation reward is not a renaming of the training objective. The central claim therefore rests on the empirical transfer properties of the learned scorer rather than on definitional equivalence.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, invented entities, or additional axioms beyond the core domain assumption stated in the text.

axioms (1)
  • domain assumption The VLM can be trained to recover temporal order from visual semantics using GRPO on shuffled sequences
    Described as the key mechanism that 'forces the model to recover temporal ordering from visual semantics rather than trivial time cues'.

pith-pipeline@v0.9.1-grok · 5748 in / 1304 out tokens · 39718 ms · 2026-07-03T17:40:41.242646+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 11 canonical work pages · 4 internal anchors

  1. [1]

    Qwen2.5-VL Technical Report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang 0028, Shijie Wang 0007, Jun Tang 0008, Humen Zhong, Yuanzhi Zhu, Ming-Hsuan Yang 0001, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang 0003, Haiyang Xu 0001, and Junyang Li...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2502.13923 2025

  2. [2]

    Enhancing vision-language model training with reinforcement learning in synthetic worlds for real-world success.CoRR, abs/2508.04280, 2025

    George Bredis, Stanislav Dereka, Viacheslav Sinii, Ruslan Rakhimov, and Daniil Gavrilov. Enhancing vision-language model training with reinforcement learning in synthetic worlds for real-world success.CoRR, abs/2508.04280, 2025. doi: 10.48550/arxiv.2508.04280. URL https://doi.org/10.48550/arxiv.2508.04280

    work page doi:10.48550/arxiv.2508.04280 2025

  3. [3]

    Brown, Wonjoon Goo, Prabhat Nagarajan, and Scott Niekum

    Daniel S. Brown, Wonjoon Goo, Prabhat Nagarajan, and Scott Niekum. Extrapolating beyond suboptimal demonstrations via inverse reinforcement learning from observations.ICML, pages 783–792, 2019. URLhttp://proceedings.mlr.press/v97/brown19a.html

    2019

  4. [4]

    Brown, Wonjoon Goo, and Scott Niekum

    Daniel S. Brown, Wonjoon Goo, and Scott Niekum. Better-than-demonstrator imitation learning via automatically-ranked demonstrations.CoRL, pages 330–359, 2019. URL http://procee dings.mlr.press/v100/brown20a.html

    2019

  5. [5]

    On the statistical efficiency of reward-free exploration in non-linear rl

    Jinglin Chen, Aditya Modi, Akshay Krishnamurthy, Nan Jiang, and Alekh Agarwal. On the statistical efficiency of reward-free exploration in non-linear rl. In S. Koyejo, S. Mo- hamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Information Processing Systems, volume 35, pages 20960–20973. Curran Associates, Inc., 2022. URL https:/...

    2022

  6. [6]

    Offline reinforcement learning for llm multi-step reasoning

    Kevin Chen, Marco F. Cusumano-Towner, Brody Huval, Aleksei Petrenko, Jackson Hamburger, Vladlen Koltun, and Philipp Kr¨ahenb¨uhl. Reinforcement learning for long-horizon interactive llm agents.CoRR, abs/2502.01600, 2025. doi: 10.48550/arxiv.2502.01600. URL https: //doi.org/10.48550/arxiv.2502.01600

    work page doi:10.48550/arxiv.2502.01600 2025

  7. [7]

    Gymnasium robotics, 2024

    Rodrigo de Lazcano, Kallinteris Andreas, Jun Jet Tai, Seungjae Ryan Lee, and Jordan Terry. Gymnasium robotics, 2024. URL http://github.com/Farama-Foundation/Gymnasiu m-Robotics

    2024

  8. [8]

    A new era of intelligence with gemini 3., 2025

    Google. A new era of intelligence with gemini 3., 2025. URL https://deepmind.google/ models/gemini/pro/

    2025

  9. [9]

    Zee- shan Zia, and Quoc-Huy Tran

    Sanjay Haresh, Sateesh Kumar, Huseyin Coskun, Shahram Najam Syed, Andrey Konin, M. Zee- shan Zia, and Quoc-Huy Tran. Learning by aligning videos in time.CVPR, pages 5548–5558, June 2021. doi: 10.1109/cvpr46437.2021.00550. URL https://doi.org/10.1109/cvpr46 437.2021.00550

    work page doi:10.1109/cvpr46437.2021.00550 2021

  10. [10]

    Generative adversarial imitation learning

    Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. InProceedings of the 30th International Conference on Neural Information Processing Systems, NIPS’16, page 4572–4580, Red Hook, NY , USA, 2016. Curran Associates Inc. ISBN 9781510838819

    2016

  11. [11]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. URL https://openreview.n et/forum?id=nZeVKeeFYf9

    2022

  12. [12]

    Imitation learning from a single temporally misaligned video

    William Huey, Huaxiaoyue Wang, Anne Wu, Yoav Artzi, and Sanjiban Choudhury. Imitation learning from a single temporally misaligned video. InForty-second International Conference on Machine Learning, 2025. URLhttps://openreview.net/forum?id=YV05KZt7v2

    2025

  13. [13]

    Kuo-Han Hung, Pang-Chi Lo, Jia-Fong Yeh, Han-Yuan Hsu, Yi-Ting Chen, and Winston H. Hsu. Victor: Learning hierarchical vision-instruction correlation rewards for long-horizon manipulation.CoRR, abs/2405.16545, 2024. doi: 10.48550/arxiv.2405.16545. URL https://doi.org/10.48550/arxiv.2405.16545. 10

    work page doi:10.48550/arxiv.2405.16545 2024

  14. [14]

    Reward-free exploration for reinforcement learning

    Chi Jin, Akshay Krishnamurthy, Max Simchowitz, and Tiancheng Yu. Reward-free exploration for reinforcement learning. In Hal Daum ´e III and Aarti Singh, editors,Proceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pages 4870–4879. PMLR, 13–18 Jul 2020. URL https://proceedings. mlr.pre...

    2020

  15. [15]

    Adaptive reward-free exploration

    Emilie Kaufmann, Pierre M ´enard, Omar Darwiche Domingues, Anders Jonsson, Edouard Leurent, and Michal Valko. Adaptive reward-free exploration. In Vitaly Feldman, Katrina Ligett, and Sivan Sabato, editors,Proceedings of the 32nd International Conference on Algorithmic Learning Theory, volume 132 ofProceedings of Machine Learning Research, pages 865–891. P...

    2021

  16. [16]

    Vision language models are in-context value learners

    Yecheng Jason Ma, Joey Hejna, Chuyuan Fu, Dhruv Shah, Jacky Liang, Zhuo Xu, Sean Kirmani, Peng Xu, Danny Driess, Ted Xiao, Osbert Bastani, Dinesh Jayaraman, Wenhao Yu, Tingnan Zhang, Dorsa Sadigh, and Fei Xia. Vision language models are in-context value learners. InThe Thirteenth International Conference on Learning Representations, 2025. URL https://open...

    2025

  17. [17]

    Zentner, Ryan Julian, J K Terry, Isaac Woungang, Nariman Farsad, and Pablo Samuel Castro

    Reginald McLean, Evangelos Chatzaroulas, Luc McCutcheon, Frank R ¨oder, Tianhe Yu, Zhan- peng He, K.R. Zentner, Ryan Julian, J K Terry, Isaac Woungang, Nariman Farsad, and Pablo Samuel Castro. Meta-world+: An improved, standardized, RL benchmark. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Bench- marks Track,...

    2025

  18. [18]

    Efros, and Trevor Darrell

    Deepak Pathak, Pulkit Agrawal, Alexei A. Efros, and Trevor Darrell. Curiosity-driven explo- ration by self-supervised prediction. InProceedings of the 34th International Conference on Machine Learning - Volume 70, ICML’17, page 2778–2787. JMLR.org, 2017

    2017

  19. [19]

    Vision- language models are zero-shot reward models for reinforcement learning

    Juan Rocamonde, Victoriano Montesinos, Elvis Nava, Ethan Perez, and David Lindner. Vision- language models are zero-shot reward models for reinforcement learning. InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.n et/forum?id=N0I2RtD8je

    2024

  20. [20]

    Proximal Policy Optimization Algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.CoRR, abs/1707.06347, 2017. URL http://dblp.uni-tri er.de/db/journals/corr/corr1707.html#SchulmanWDRK17

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  21. [22]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.CoRR, abs/2402.03300, 2024. doi: 10.48550/arxiv.2402.03300. URL https://doi.org/10.48550/arxiv.2402.03300

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2402.03300 2024

  22. [23]

    Coin: A large-scale dataset for comprehensive instructional video analysis

    Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  23. [24]

    Generative adversarial imitation from observa- tion

    Faraz Torabi, Garrett Warnell, and Peter Stone. Generative adversarial imitation from observa- tion. InImitation, Intent, and Interaction (I3) Workshop at ICML 2019, June 2019

    2019

  24. [25]

    Time your rewards: Learning temporally consistent rewards from a single video demonstration

    Huaxiaoyue Wang, William Huey, Anne Wu, Yoav Artzi, and Sanjiban Choudhury. Time your rewards: Learning temporally consistent rewards from a single video demonstration. In CoRL 2024 Workshop on Whole-body Control and Bimanual Manipulation: Applications in Humanoids and Beyond, 2024. URLhttps://openreview.net/forum?id=gsgkiuv9BS

    2024

  25. [26]

    Rank2reward: Learning shaped reward functions from passive video.ICRA, 2024

    Daniel Yang, Davin Tjia, Jacob Berg, Dima Damen, Pulkit Agrawal, and Abhishek Gupta. Rank2reward: Learning shaped reward functions from passive video.ICRA, 2024. 11

    2024

  26. [27]

    Mastering visual continuous control: Improved data-augmented reinforcement learning, 2021

    Denis Yarats, Rob Fergus, Alessandro Lazaric, and Lerrel Pinto. Mastering visual continuous control: Improved data-augmented reinforcement learning, 2021. URL https://arxiv.org/ abs/2107.09645

    work page arXiv 2021

  27. [28]

    Xirl: Cross-embodiment inverse reinforcement learning.Conference on Robot Learning (CoRL), 2021

    Kevin Zakka, Andy Zeng, Pete Florence, Jonathan Tompson, Jeannette Bohg, and Debidatta Dwibedi. Xirl: Cross-embodiment inverse reinforcement learning.Conference on Robot Learning (CoRL), 2021

    2021

  28. [29]

    A vision-language-action-critic model for robotic real-world reinforcement learning.CoRR, abs/2509.15937, 2025

    Shaopeng Zhai, Qi Zhang, Tianyi Zhang, Fuxian Huang, Haoran Zhang, Ming Zhou, Shengzhe Zhang, Litao Liu, Sixu Lin, and Jiangmiao Pang. A vision-language-action-critic model for robotic real-world reinforcement learning.CoRR, abs/2509.15937, 2025. doi: 10.48550/arxiv.2 509.15937. URLhttps://doi.org/10.48550/arxiv.2509.15937

    work page doi:10.48550/arxiv.2 2025

  29. [30]

    Fine-tuning large vision-language models as decision-making agents via reinforcement learning.CoRR, abs/2405.10292, 2024

    Yuexiang Zhai, Hao Bai, Zipeng Lin, Jiayi Pan 0002, Shengbang Tong, Yifei Zhou, Alane Suhr, Saining Xie, Yann LeCun, Yi Ma 0001, and Sergey Levine. Fine-tuning large vision-language models as decision-making agents via reinforcement learning.CoRR, abs/2405.10292, 2024. doi: 10.48550/arxiv.2405.10292. URLhttps://doi.org/10.48550/arxiv.2405.10292

    work page doi:10.48550/arxiv.2405.10292 2024

  30. [31]

    VideoGameBench: Can Vision-Language Models complete popular video games?

    Alex L. Zhang, Thomas L. Griffiths, Karthik R. Narasimhan, and Ofir Press. Videogamebench: Can vision-language models complete popular video games?, 2025. URL https://arxiv.or g/abs/2505.18134. 12 0 200 400 600 800 1000 1200 1400 Training Steps 0.0 0.2 0.4 0.6 0.8 1.0Success Rate LOOP w/ refresh LOOP w/o refresh Starting point refresh 0 200 400 600 800 10...

    work page internal anchor Pith review Pith/arXiv arXiv 2025