Robotics · arXiv:2610.16641 v1 · submitted 12 Oct 2026 · 30 references
What Limits Cross-Task Generalization in World Action Models?
The paper says
“Cross-task generalization remains difficult for world action models (WAMs), even when their predicted futures depict plausible manipulation.”
more of the abstract
This gap can be attributed directly: we make this gap directly testable by decoupling a WAM into a world model and an inverse-dynamics model (IDM), then holding the IDM fixed while replacing the source of future images. We introduce Align Once, Evaluate Everywhere (AoE), a camera-space inverse-dynamics model pretrained on large-scale robot data, which provides this recovery interface for different future sources across tasks and environments. Through it, we find that actions recovered from recorded futures of unseen tasks execute almost as well as replayed demonstrations, whereas futures predicted by world models collapse across benchmarks. The failures are not generic: they concentrate on the gripper and interaction geometry while generation capacity is spent on task-irrelevant background. This diagnosis suggests that supervision should target exactly these regions. We therefore introduce the vision-to-action (VITA) loss, which passes generated futures through the frozen AoE and penalizes discrepancies in the recovered actions, redirecting generation toward action-relevant detail. Across two world models and three manipulation benchmarks, this simple change raises the execution success of generated futures from near zero to 47.0%. The same loss also transfers to WAMs and yields consistent gains in native control.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 14 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
The passes disagreed. Reader: holds up. Skeptic: agrees with the reader. Referee: major revision.
Solid paper: VITA makes world-model futures executable, with extensive experiments and honest limitations; the circularity concern is partially addressed but the cleanest resolution (ground-truth oracle simulation) is missing, though not fatal. This line is wrong
2 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 30 references say.
The paper says, at §4.2, Appendix H
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.65):
“This makes it difficult to directly assess how the predicted visual future itself supports execution on unseen tasks.”
Pith says: Objection 1 of 2 · major
The central claim is that VITA supervision makes generated futures truly more executable. The main evaluation uses AoE as both the VITA teacher and the execution evaluator. Appendix H partially addresses this by showing that VITA-supervised futures also improve when evaluated by J-IDM and DreamGen-IDM.
more of objection 1
However, all three IDMs are learned from data and may share biases in handling unrealistic generated images (especially at the 0.3% baseline). The simulation environments (LIBERO, RoboTwin) provide ground-truth object and end-effector poses; a perfect motion planner using these poses would yield a success rate that reflects intrinsic executability without any learned IDM bias. The paper does not include such an experiment. Without it, there remains a possibility that VITA-supervised futures improve primarily because they become easier to decode for learned IDMs (which share training-data distribution with AoE), not because they encode more accurate physical interaction information. I request either (a) a ground-truth oracle execution experiment using simulator state to directly measure whether VITA-supervised futures encode better controllability, or (b) a detailed discussion explaining why such an experiment is infeasible and why the cross-IDM evidence is sufficient.
The paper says, at §3.3, Eq. (10), Appendix E
The review did not quote this passage.
Pith says: Objection 2 of 2 · major
The local action-error bound in Eq. (10) depends on the assumption that 'AoE captures the feature directions used by the action head' (Eq. (31)). This is stated as a condition, but the paper does not provide empirical verification that this condition holds for the actual networks used (e.g., does AoE's Jacobian span the relevant directions of the LingBot-VA action head?).
more of objection 2
While the bound is presented as an analytical insight, the practical relevance hinges on this transfer condition. I suggest either providing a direct measurement (e.g., comparing the subspace of AoE's Jacobian to the action head's sensitivity) or adding a caveat that the bound is interpretative and may not hold tightly in practice.
5 minor objections
The paper says, at Abstract
Pith says: minor objection 1 of 5
The abstract states 'actions recovered from recorded futures of unseen tasks execute almost as well as replayed demonstrations.' This is supported by Table 1, but the wording 'almost as well' could be quantified: on LIBERO, recorded futures + AoE achieve 57.0% vs. replayed actions at 72.3%, a gap of ~15 points. Consider providing the exact numbers or rephrasing to avoid overstatement.
The paper says, at Table 3
Pith says: minor objection 2 of 5
The green/red color coding for improvements/degradations is helpful, but the table caption does not explain the color scheme. Please add a brief note in the caption or below the table.
show 3 more
The paper says, at §4.3, Figure 3
Pith says: minor objection 3 of 5
The x-axis of Figure 3 is labeled 'Pretraining data (hours)'. The three subplots share the same x-axis but have different y-axes. The reader should be able to see the scale more clearly; consider adding an explicit tick for the 500-hour point.
The paper says, at Appendix E, Eq. (32)
Pith says: minor objection 4 of 5
The bound includes a term C_h ∥h - h*∥_2^2. It would be helpful to briefly comment on the magnitude of C_h, e.g., whether it is expected to be small relative to the other terms based on the Lipschitz constants of the networks.
The paper says, at General
Pith says: minor objection 5 of 5
The paper uses the notation 'WAM' to mean 'world action model' and 'WM' for 'world model'. Ensure that this distinction is consistently applied; in a few places (e.g., line after Eq. (4)), the two terms are used interchangeably, which could confuse readers.
What would settle it
Use the simulated LIBERO and RoboTwin environments to extract ground-truth end-effector and object poses from both recorded and generated future video frames. For each generated future sequence, use these ground-truth poses to produce a perfect action sequence via a kinematic motion planner (e.g., inverse kinematics with joint limits). Execute these actions in the simulator and compare success rates between VITA-supervised and baseline world models, and against the recorded-future oracle. If VITA-supervised futures achieve significantly higher success with ground-truth pose-based execution (approaching the recorded-future oracle), the concern is largely resolved. If the gap remains large or matches the baseline, then the reported gains reflect alignment to learned IDM biases rather than genuinely improved visual executability.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 14 Oct 2026
- Reader holds up
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 2 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
The paper does two things well. First, it isolates the bottleneck in cross-task world model generalization: the generated futures look plausible but fail because gripper and interaction geometry are wrong, not because the video is blurry. Second, it introduces VITA, a simple loss that uses a frozen pretrained inverse dynamics model (AoE) to penalize action-space discrepancies between generated and recorded futures. That is a clean, reusable idea, and the results are striking — from near 0% to ~47% execution on LIBERO, with consistent gains on RoboTwin and LIBERO-Plus.
The experiments are thorough: two world models, two world action models, multiple benchmarks, ablations of the matching loss and GT loss, scaling studies, and teacher-independent evaluation (Appendix H). The mathematical analysis in Appendix E is careful, and the paper is transparent about limitations (Section 6: camera occlusions, safety concerns, residual gap to recorded replay).
The main soft spot is the circularity of using AoE as both the VITA supervisor and the evaluator. The authors mitigate this by testing VITA-supervised futures with different IDMs (J-IDM, DreamGen-IDM) and showing persistent improvement. That helps, but it does not fully settle the stress-test concern: all learned IDMs may share biases from training data distributions, especially on OOD generated images. A ground-truth oracle experiment in simulation (using perfect state-based execution) would be the cleanest check, and the paper does not include it. That said, the teacher-independent results are strong enough that I do not think this is a fatal flaw — the paper reports a genuine improvement, and the residual uncertainty is well within the bounds of what a good reviewer discussion can resolve.
A minor point: the AoE checkpoint is promised for future release but not yet available, which limits immediate reproduction. The paper is written clearly and honestly, and the architecture and loss are described in enough detail for reimplementation.
Bottom line: This is a solid paper with a useful method and honest evaluation. It deserves peer review and will likely be accepted after addressing the circularity concern more directly (e.g., by running the oracle simulation or clarifying why it is infeasible). I would bring it to my reading group, and if I worked on world models for manipulation, I would cite it.
The strongest objection
The central claim is that VITA supervision makes generated futures executable. The paper provides teacher-independent evaluations (Appendix H) that show VITA-trained futures also improve when evaluated by J-IDM and DreamGen-IDM, arguing against pure overfitting to AoE. However, all these IDMs are learned from data and may share limitations in handling out-of-distribution visual features, especially on the highly unrealistic generated images from the baseline (0.3% success). The ground-truth simulation environment (LIBERO, RoboTwin) can directly compute task success from state, bypassing any learned IDM. If a perfect motion planner (using ground-truth end-effector and object poses) is used to execute the generated future videos, the measured success rate would reflect intrinsic executability, free from learned IDM biases. The paper does not include such an experiment. Without it, there remains a possibility that VITA-supervised futures improve execution primarily because they become easier to decode for learned IDMs (which share training data distributions with AoE), not because they contain more accurate physical interaction information. This is load-bearing because it touches the core interpretation of whether VITA truly improves the world model's understanding of manipulable geometry or merely adapts it to a specific action-interface family.
The paper's central claim is strongly supported by extensive experiments, including teacher-independent evaluations and scaling studies. The load-bearing concern is not that the results are wrong, but that the mechanism of improvement might be partially specific to learned IDM interfaces rather than a general improvement in executability. The reader's weakest assumption focused on AoE's reliability on OOD generated images; my concern is more specific: even if AoE's predictions are reliable, the improvement might be due to alignment of the generated images to the AoE's action-interface biases, which could be a narrower improvement than claimed. However, the paper already shows that VITA supervision improves performance with two different world models and three benchmarks, and the teacher-independent evaluation partially addresses this. The proposed concrete test (ground-truth oracle execution) would provide definitive evidence that the futures encode genuinely better spatial interaction geometry. I agree with the reader's ACCEPT verdict because the existing evidence is compelling and the limitations are honestly stated. The concern does not invalidate the contribution, but it identifies a missing control that would strengthen the interpretation. Therefore, the verdict remains UNCHANGED: the paper still merits acceptance based on the presented evidence.
The referee's report
If the results hold, this paper makes three significant contributions. (1) It introduces AoE, a camera-space inverse-dynamics model pretrained on large-scale robot data that provides a practical zero-shot action-recovery interface for evaluating and supervising world models across unseen tasks. (2) It proposes the VITA loss, a simple yet effective method that uses a frozen IDM to inject action-grounded supervision into world-model training, substantially closing the gap between generated and recorded futures. (3) The extensive experimental validation—two world models, two WAMs, three benchmarks, multiple ablations, scaling studies, and teacher-independent controls (Appendix H)—is a strength of the paper. The work also includes formal analysis of the VITA loss (Appendix E). The improvements are large (e.g., from 0.4% to 47% on LIBERO) and are replicated across different evaluator IDMs, suggesting that the effect is not merely superficial alignment to AoE.
Pith's reading of the paper (a machine's summary, not the authors')
This paper identifies and addresses a key bottleneck in world action models: futures predicted by world models look plausible but fail when used to generate executable actions on unseen tasks. By introducing Align Once, Evaluate Everywhere (AoE), a camera-space inverse-dynamics model pretrained on 9,637 hours of robot data, the authors create a fixed interface to evaluate and supervise world models through their downstream action recoverability. Using AoE, they show that failures concentrate on gripper and interaction geometry, while generation capacity is wasted on background. They then propose the vision-to-action (VITA) loss, which passes generated futures through the frozen AoE and penalizes discrepancies in the recovered actions, redirecting generation toward action-relevant detail. This simple change raises cross-task execution success of generated futures from near zero to 47.0–51.7% across benchmarks and also improves native action prediction in world action models, especially under distribution shifts.
What carries the argument. The paper's central mechanism is the combination of Align Once, Evaluate Everywhere (AoE) and the Vision-to-Action (VITA) loss. AoE is a pretrained inverse-dynamics model that uses a camera-space action representation—projected end-effector coordinates, depth change, rotation, and gripper command—to recover executable actions from image pairs in a way that transfers zero-shot across environments without relearning coordinate conventions. The VITA loss passes future images predicted by a world model through this frozen AoE and penalizes both the difference between the recovered action and the action recovered from the corresponding recorded future (matching loss) and the difference between the recovered action and the ground-truth action that produced the transition (GT loss). This provides a differentiable, action-grounded training signal that is particularly sensitive to errors in the gripper and interaction region, as measured by AoE's input derivatives. Gradients propagate through the generated image branch while AoE remains frozen, updating the world model to focus on the visual details that matter for downstream execution.
Core claim. The paper's central discovery is that the cross-task generalization failure of world action models stems not from generic prediction quality but from a loss of action-relevant detail in generated futures, particularly in gripper and interaction geometry. By decoupling the world model from action recovery via a frozen, camera-space inverse-dynamics model (AoE), the authors provide the first direct execution-based evaluation of world model futures and show that recorded images support near-demonstration-level success while generated futures collapse to near zero. They further show that supervising world model training with an action-matching loss (VITA) that compares actions recovered from generated and recorded futures, using the same frozen AoE as teacher, forces the world model to preserve the visual details needed for physical execution. This supervision raises execution success from ~0.4% to 47.0% on LIBERO and from ~1.3% to 40.4% on RoboTwin Hard, and it also improves the native action prediction of world action models (e.g., LingBot-VA from 35.1% to 47.2% on LIBERO-Plus) by shaping the shared representations used for control, even when the model never explicitly renders images at inference.
The premise it rests on. The frozen AoE teacher provides sufficiently accurate and stable action-flow predictions on out-of-distribution generated images to supervise world model training, even though those generated images are initially highly unrealistic and the teacher's accuracy on such images is not directly verified.
If the paper is right (Pith's inferences, not the authors')
World models can now be evaluated by their downstream executability rather than visual fidelity alone; the AoE interface provides a standardized, deployment-relevant metric.
The VITA loss provides a general-purpose, action-grounded supervision that improves both pure world models and integrated world action models, with particularly large gains under distribution shifts.
Larger-scale pretraining of the inverse-dynamics teacher monotonically improves both the quality of evaluation and the strength of VITA supervision, suggesting that further scaling of AoE would yield continued downstream gains.
Since FastWAM benefits from VITA training even though it never renders images at inference, the supervision works by shaping shared representations, implying that any architecture with a visual prediction path can be improved by action-consistent training without requiring test-time video generation.
The camera-space action representation enables zero-shot transfer across environments without adapting the IDM, reducing the cost of deploying world models in new robot setups.
The VITA loss implies that the primary bottleneck in world model generalization is not global scene consistency but fine-grained manipulation geometry; future world model architectures might explicitly allocate capacity to end-effector and contact regions, perhaps through learned attention masks or separate decoder branches.
The AoE teacher’s sensitivity to gripper motion suggests that inverse dynamics models can serve as an automatic reward signal for training world models in open-ended environments, potentially reducing reliance on dense human demonstrations.
Since VITA-supervised futures remain below recorded-video replay in execution success, combining action-grounded supervision with explicit 3D structure prediction or closed-loop correction during execution might close the remaining gap.
The failure of VITA under severe occlusion (gripper leaving camera view) is acknowledged; a natural extension would be to incorporate multiple camera views or active camera control to maintain visibility of the end-effector during training and evaluation.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Cross-task generalization remains difficult for world action models (WAMs), even when their predicted futures depict plausible manipulation. This gap can be attributed directly: we make this gap directly testable by decoupling a WAM into a world model and an inverse-dynamics model (IDM), then holding the IDM fixed while replacing the source of future images. We introduce Align Once, Evaluate Everywhere (AoE), a camera-space inverse-dynamics model pretrained on large-scale robot data, which provides this recovery interface for different future sources across tasks and environments. Through it, we find that actions recovered from recorded futures of unseen tasks execute almost as well as replayed demonstrations, whereas futures predicted by world models collapse across benchmarks. The failures are not generic: they concentrate on the gripper and interaction geometry while generation capacity is spent on task-irrelevant background. This diagnosis suggests that supervision should target exactly these regions. We therefore introduce the vision-to-action (VITA) loss, which passes generated futures through the frozen AoE and penalizes discrepancies in the recovered actions, redirecting generation toward action-relevant detail. Across two world models and three manipulation benchmarks, this simple change raises the execution success of generated futures from near zero to 47.0%. The same loss also transfers to WAMs and yields consistent gains in native control.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-14
The authors' main claim, as the classifier picked it out 0.35
“The failures are not generic: they concentrate on the gripper and interaction geometry while generation capacity is spent on task-irrelevant background.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.29 · human participants 0.02 · controlled comparison 0.87 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.82 · mentions code or data availability 0.04
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Cross-task generalization remains difficult for world action models (WAMs), even when their predicted futures depict plausible manipulation. | background 0.79 | 0.08 | 0.06 | asserted 1.00 |
| This gap can be attributed directly: we make this gap directly testable by decoupling a WAM into a world model and an inverse-dynamics model (IDM), then holding the IDM fixed while replacing the source of future images. | method 0.68 | 0.84 | 0.08 | asserted 0.86 |
| We introduce Align Once, Evaluate Everywhere (AoE), a camera-space inverse-dynamics model pretrained on large-scale robot data, which provides this recovery interface for different future sources across tasks and environments. | method 0.66 | 0.89 | 0.05 | asserted 1.00 |
| Through it, we find that actions recovered from recorded futures of unseen tasks execute almost as well as replayed demonstrations, whereas futures predicted by world models collapse across benchmarks. | result 1.00 | 0.96 | 0.53 | asserted 1.00 |
| The failures are not generic: they concentrate on the gripper and interaction geometry while generation capacity is spent on task-irrelevant background. | result 0.53 | 0.90 | 0.20 | asserted 1.00 |
| This diagnosis suggests that supervision should target exactly these regions. | interpretation 0.95 | 0.49 | 0.24 | hedged possibility 0.98 |
| We therefore introduce the vision-to-action (VITA) loss, which passes generated futures through the frozen AoE and penalizes discrepancies in the recovered actions, redirecting generation toward action-relevant detail. | method 0.74 | 0.93 | 0.15 | asserted 0.99 |
| Across two world models and three manipulation benchmarks, this simple change raises the execution success of generated futures from near zero to 47.0%. | result 1.00 | 0.96 | 0.39 | asserted 1.00 |
| The same loss also transfers to WAMs and yields consistent gains in native control. | result 1.00 | 0.94 | 0.18 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 0.98 · evidence from observation or experiment 0.98
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-14: 1 request, 6,929 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (10)
Figures are the authors', reproduced from the paper.
References (30)
- [1] Cosmos World Foundation Model Platform for Physical. arXiv preprint arXiv:2501.03575 , year = on PithPith's reviewarXiv
- [2] Causal World Modeling for Robot Control , author =. arXiv preprint arXiv:2601.21998 , year = on PithPith's reviewarXiv
- [3] Yuan, Tianyuan and Dong, Zibin and Liu, Yicheng and Zhao, Hang , journal =. 2026 , url = on Pith
- [4] Shang, Yu and Li, Zhuohang and Ma, Yiding and Su, Weikang and Jin, Xin and Wang, Ziyou and Jin, Lei and Zhang, Xin and Tang, Yinzhou and Su, Haisheng and Gao, Chen and Wu, Wei and Liu, Xihui and Shah, Dhruv and Zhang, Zhaoxiang and Chen, Zhibo and Zhu, Jun and Tian, Yonghong a... on Pith
- [5] Liu, Zeyu and Zhu, Zhangzhe and Zhang, Yang and Fan, Chenyou and Bai, Chenjia and Li, Xuelong , journal =. 2026 , url = on Pith
- [6] Co, Peterson and Hu, Sicheng and Jiao, Chunxuan and Cheng, Hongyang and Luo, Yulin and Xu, Yijie and Chen, Sixiang and Zhao, Zhongxia and Wang, Zihao and Chi, DaFeng and Liu, Peidong and Chen, YuTong and Liu, Henghua and Yuan, Zhihao and Jia, Huizhu and Zhuang, Yuzheng and Zha... on Pith
- [7] Liu, Bo and Zhu, Yifeng and Gao, Chongkai and Feng, Yihao and Liu, Qiang and Zhu, Yuke and Stone, Peter , booktitle =. 2023 , url = on Pith
- [8] Fei, Senyu and Wang, Siyin and Shi, Junhao and Dai, Zihao and Cai, Jikun and Qian, Pengfang and Ji, Li and He, Xinzhe and Zhang, Shiduo and Fei, Zhaoye and Fu, Jinlan and Gong, Jingjing and Qiu, Xipeng , booktitle =. 2026 , url = on Pith
- [9] Jang, Joel and Ye, Seonghyeon and Lin, Zongyu and Xiang, Jiannan and Bjorck, Johan and Fang, Yu and Hu, Fengyuan and Huang, Spencer and Kundalia, Kaushil and Lin, Yen-Chen and Magne, Lo. Proceedings of the 9th Conference on Robot Learning , pages =. 2025 , publisher = on Pith
- [10] Turning Video Models into Generalist Robot Policies , author =. arXiv preprint arXiv:2605.27817 , year = on PithPith's reviewarXiv
- [11] arXiv preprint arXiv:2609.05588 , year = on PithPith's reviewarXiv
- [12] Chen, Tianxing and Chen, Zanxin and Chen, Baijun and Cai, Zijian and Liu, Yibin and Li, Zixuan and Liang, Qiwei and Lin, Xianliang and Ge, Yiheng and Gu, Zhenyu and Deng, Weiliang and Guo, Yubin and Nian, Tian and Xie, Xuanbing and Chen, Qiangyu and Su, Kailun and Xu, Tianling... on Pith
- [13] DINOv3 , author =. arXiv preprint arXiv:2508.10104 , year = on PithPith's reviewarXiv
- [14] On the Continuity of Rotation Representations in Neural Networks , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2019 , url = on Pith
- [15] Flow Matching for Generative Modeling , author =. International Conference on Learning Representations , year = on Pith
- [16] Wan: Open and Advanced Large-Scale Video Generative Models , author =. arXiv preprint arXiv:2503.20314 , year = on PithPith's reviewarXiv
- [17] 2025 , howpublished = on Pith
- [18] The Unreasonable Effectiveness of Deep Features as a Perceptual Metric , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , year = on Pith
- [19] Towards Accurate Generative Models of Video: A New Metric & Challenges , author =. arXiv preprint arXiv:1812.01717 , year = on PithPith's reviewarXiv
- [20] Image Quality Assessment: From Error Visibility to Structural Similarity , author =. IEEE Transactions on Image Processing , volume =. 2004 , doi = on Pith
show 10 more
- [21] EVA: Aligning Video World Models with Executable Robot Actions via Inverse Dynamics Rewards , author =. 2026 , eprint = on Pith
- [22] Wi, Youngsun and Merwe, Mark Van der and Florence, Pete and Zeng, Andy and Fazeli, Nima , booktitle =. 2023 , editor = on Pith
- [23] Spatial Action Maps for Mobile Manipulation , author =. 2020 , eprint =. doi:https://doi.org/10.15607/RSS.2020.XVI.035 , url = on Pithdoi:10.15607/rss.2020.xvi.035
- [24] Yupeng Zheng and Xiang Li and Songen Gu and Yuhang Zheng and Shuai Tian and Weize Li and Linbo Wang and Chaoyue Li and Qichao Zhang and Haoran Li and Zhongpu Xia and Ya-Qin Zhang and Shuicheng Yan and Dongbin Zhao , year =. 2609.04193 , archivePrefix = on PithPith's reviewarXiv
- [25] Action Images: End-to-End Policy Learning via Multiview Video Generation , author =. 2026 , eprint = on Pith
- [26] Feng Jiang and Yang Chen and Kyle Xu and Yuchen Liu and Haifeng Wang and Zhenhao Shen and Jasper Lu and Shengze Huang and Yuanfei Wang and Chen Xie and Ruihai Wu , year =. 2604.19092 , archivePrefix = on PithPith's reviewarXiv
- [27] Zeyi Liu and Shuang Li and Eric Cousineau and Siyuan Feng and Benjamin Burchfiel and Shuran Song , year =. Geometry-aware. 2507.01099 , archivePrefix = on PithPith's reviewarXiv
- [28] Weishi Mi and Yong Bao and Xiaowei Chi and Xiaozhu Ju and Zhiyuan Qin and Kuangzhi Ge and Kai Tang and Peidong Jia and Shanghang Zhang and Jian Tang , year =. 2601.18323 , archivePrefix = on PitharXiv
- [29] Robotic Manipulation by Imitating Generated Videos Without Physical Demonstrations , author =. 2026 , eprint = on Pith
- [30] Yao Feng and Chendong Xiang and Xinyi Mao and Hengkai Tan and Zuyue Zhang and Shuhe Huang and Kaiwen Zheng and Haitian Liu and Hang Su and Jun Zhu , year =. 2512.17661 , archivePrefix = on PitharXiv
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment