Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Is an object-centric representation beneficial for robotic manipulation ?

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Object-centric representations, which parse scenes into per-object slots, make robotic manipulation policies more robust to unseen distractors than global encodings, according to a new simulated benchmark.

desk verdict Useful benchmark wrapped around a confounded comparison: SAVi gets in-domain pre-training while DINO/R3M are frozen off-the-shelf, so the central claim about object-centric benefits is not established. read the letter →

arxiv 2506.19408 v1 pith:SEA2T5YA submitted 2025-06-24 cs.AI cs.RO

Alexandre Chapin (imagine) , Emmanuel Dellandrea (imagine) , Liming Chen (imagine) This is my paper
classification cs.AIcs.RO
keywords object-centricrepresentationroboticmanipulationimitationlearningbehaviorcloningslotattentionout-of-distributiongeneralizationSAVimulti-objectscenes
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 asks whether representing a scene as a set of per-object slots, rather than as a single global vector, helps a robot learn manipulation skills that generalize. To test this, the authors build a simulated multi-object benchmark with three pick-and-place tasks and heavy randomization, and compare a slot-based encoder (SAVi) against two frozen global representations (DINO and R3M) under unseen distractor colors, background colors, and distractor sizes. The paper's central claim is that object-centric representations overcome out-of-distribution shifts and solve a precise pick-and-place task that both global baselines fail entirely. The results are presented as evidence that structured per-object bottlenecks, rather than parameter scale or dynamics pre-training alone, are what makes the policy robust.

What carries the argument

SAVi (Slot-Attention for Video), an encoder that iteratively clusters visual features into a fixed set of K per-object slot vectors via slot attention, with a transformer predictor propagating slots across frames and a decoder reconstructing the video. The policy is a transformer-based observation trunk that takes the history of slot sets plus a learnable [ACT] token; the token's output feeds a Gaussian-mixture action head. The slot bottleneck is what isolates objects from background, and the paper argues that this isolation is the mechanism behind the generalization results.

What would settle it

Give a holistic (non-slot) encoder the same in-domain reconstruction pre-training on the RoboShape data, freeze it, and run the Pick cube evaluation under the L1/L2/L3 shifts; if its success rate reaches SAVi's 0.44-0.56 range, the claim that object-centric structure is responsible collapses.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that an object-centric video representation, SAVi, enables a behavior-cloning policy to succeed at grasping and placing a target cube under three types of distribution shift, while the global representations DINO and R3M achieve zero success on that task. In the Push cube task, R3M leads in-domain (0.88 vs 0.69 for SAVi) but loses 60% of its success rate under unseen distractor colors, whereas SAVi drops only 14%. The paper attributes this robustness to the slot mechanism, which separates objects from the background and from distractors, and reports that all models fail to place the cube in a bin due to single-view depth ambiguity.

Load-bearing premise

The paper assumes the difference between SAVi and the global baselines comes from the slot structure, but SAVi was pre-trained on the same in-domain data used for policy learning while DINO and R3M were frozen from unrelated data, so a holistic encoder given the same in-domain pre-training could potentially match SAVi.

Editorial extensions

If this is right

  • If the claim is right, object-centric encoders are a viable backbone for vision-based behavior cloning in multi-object scenes, without large-scale out-of-domain pre-training.
  • Robotic policies built on slots should degrade more gracefully than global encoders when distractor colors, background colors, or object sizes shift away from training.
  • The single-view failure of Task 3 implies that object-centric perception alone does not remove depth ambiguity; adding camera views or proprioceptive cues would be a next step.
  • Success on Pick cube suggests slot representations carry enough spatial information for relative placement, not just classification or segmentation.
  • The small parameter count of SAVi (approximately 25 times fewer) indicates that structure may substitute for scale in sample-limited manipulation learning.

Reading between the lines

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

  • The paper's comparison leaves a confound open: SAVi is pre-trained with reconstruction on the in-domain RoboShape data, while DINO and R3M are frozen from unrelated corpora, so the results do not fully isolate representation structure from training-domain overlap.
  • If the object-centric benefit is real, a decisive test is to give DINO or R3M the same in-domain reconstruction pre-training and re-run the benchmark; unless that closes the gap, the slot-mechanism account is supported.
  • The observed robustness pattern suggests that object-centric world models, which predict slot dynamics, could extend these gains beyond fixed-horizon behavior cloning to long-horizon planning.
  • The Task 3 failure mode, where the robot perceives the cube as already inside the bin, points to a depth-estimation bottleneck that a multi-view or depth-sensing variant could resolve, which is a testable design prediction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper asks whether an object-centric visual representation benefits imitation learning for robotic manipulation. The authors propose a framework that combines a SAVi object-centric video encoder with a transformer-based behavior cloning policy, introduce a simulated benchmark (RoboShape) with three manipulation tasks and three types of generalization shifts (unseen distractor colors, background colors, and distractor sizes), and compare SAVi against two frozen global representations, DINO and R3M. They report that SAVi succeeds on two tasks and degrades only modestly under distribution shift, while DINO fails on all tasks and R3M either fails or degrades sharply. The central claim is that object-centric structure, rather than pre-training scale or temporal dynamics, explains the observed success and robustness. The paper also identifies single-view ambiguity as the likely cause of total failure on the third task.

Significance. If the central claim were established, the paper would provide valuable evidence for the practical benefit of object-centric representations in robot manipulation, and the RoboShape benchmark could be a useful resource for the community. The reporting discipline is a strength: 100 rollouts per scenario, three repeats, and means with standard deviations are given. The paper also makes a fair observation that global representations can fail catastrophically when scene structure is complex. However, the current experimental protocol cannot support the causal attribution of the performance difference to object-centric structure, because SAVi is the only method that is pre-trained on in-domain data with a reconstruction objective. This confound is load-bearing for the paper's main conclusion, so the significance is currently conditional on a controlled re-evaluation.

major comments (4)
  1. [Section 3.2 (Training) and Section 4.1] The comparison is systematically confounded: SAVi's encoder is pre-trained with an MSE reconstruction objective on in-domain data that are exactly the same data used for the behavior learning phase ('the model is trained on in-domain data (the same data used for the behavior learning phase)'), whereas DINO and R3M are frozen off-the-shelf encoders from unrelated pre-training corpora (Section 4.1, 'All models are frozen during policy learning'). Therefore the Pick cube result in Table 2 (SAVi 0.56 in-domain, 0.44-0.51 under shifts, versus 0.0 for both baselines) could be explained by SAVi's familiarity with the simulator's rendering, background palette, and object geometry rather than by slot decomposition. The paper needs a holistic encoder trained on the same RoboShape data with the same reconstruction objective and frozen under the same protocol; without that baseline, the central claim that object-centric structure is beneficial is not supported.
  2. [Section 5 (Conclusion)] The conclusion states that 'object-centric methods can mitigate the effects of out-of-distribution scenarios, such as handling new distractors.' This claim is not supported by the current protocol because the robustness advantage of SAVi in Tables 1 and 2 may reflect in-domain pre-training rather than any property of the object-centric bottleneck. The paper should either provide the missing controlled baseline or explicitly restrict the claim to 'the tested SAVi model with in-domain reconstruction pre-training' rather than 'object-centric methods.' This is a load-bearing issue for the title question and the abstract.
  3. [Section 4.2 (Task 3)] Task 3 (Place cube in bin) yields zero success for all models, and the authors attribute the failure to single-view ambiguity, illustrated in Figure 6 ('the model appears to perceive the cube as already inside the box'). This means that the proposed benchmark's most complex task provides no evidence for the claimed advantage of object-centric representations. The claim that object-centric methods 'help overcome these challenges' in complex manipulation scenarios is therefore supported by only two tasks, one of which is confounded as described above. The authors should either add a multi-view setup or substantially temper the scope of the claim.
  4. [Section 4.2 (Quantitative Results)] The paper does not report statistical significance tests, so differences such as the Push cube L1 result (R3M 0.26 +/- 0.06 vs SAVi 0.55 +/- 0.02) are presented as robust trends without a formal comparison. Given the large variance in some cells and the small number of repeats, adding a significance test or confidence intervals would strengthen the claims. This comment is secondary to the confound issue but relevant to the interpretation of the generalization results.
minor comments (5)
  1. [Abstract and Section 2] There are repeated typos: 'hollistic' should be 'holistic', 'conises' in Section 3.1 should be 'consists', and 'realease' in Section 4.2 should be 'release'.
  2. [Section 3.2] Key hyperparameters of SAVi are not reported: the number of slots K, the slot dimension D, the history length H, the number of GMM components, and the exact architecture of the observation trunk. Without these, the experiments are difficult to reproduce.
  3. [Tables 1 and 2] In Tables 1 and 2, the DINO column is listed as '0.' with no standard deviation. This is presumably because all three repeats are zero, but the formatting should be clarified, for example by writing '0.00 ± 0.00'.
  4. [Figure 6] The caption of Figure 6 says 'Model reconstruction', but it is not clear whether this reconstruction is produced by the SAVi decoder from slots or by the policy's predicted next slot. Please clarify the source of the reconstruction.
  5. [Limitations and future works] The limitations paragraph acknowledges the simulation-to-real gap and the simplicity of the tasks, but it does not acknowledge the pre-training protocol asymmetry between SAVi and the baselines, which is a threat to the validity of the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the evaluation is an empirical benchmark against external baselines, and the in-domain pre-training asymmetry is a soundness/protocol confound rather than a circular derivation.

full rationale

This paper contains no derivation chain of the kind that could be circular: it builds a transformer-based behavior-cloning policy on top of a SAVi encoder and compares it with DINO and R3M encoders on simulated RoboShape tasks. The conclusion that 'object-centric methods can mitigate the effects of out-of-distribution scenarios' is supported by success-rate tables, not by equations that define the conclusion into the model. No parameter is fitted to a subset and then reported as a prediction of a closely related quantity; the L1-L3 generalization shifts use unseen colors, backgrounds, and sizes, so they are not identical to the training input by construction. There are also no load-bearing self-citations: none of the cited works is by the authors (Chapin, Dellandrea, Chen), and the adopted architectures (SAVi [20], BAKU-style trunk [12]) are external prior work used as components, not as justification of the central claim. The paper's Limitations section honestly confines the claims to simulation, and that limitation does not create circularity. The main validity concern is a protocol asymmetry: SAVi is pre-trained on in-domain RoboShape data with an MSE reconstruction loss, while DINO and R3M are frozen from unrelated corpora, and parameter scales differ. That asymmetry means Table 2's Pick cube result cannot cleanly attribute SAVi's advantage to object-centric structure rather than in-domain pre-training or architecture. However, this is an experimental confound and a soundness/fairness issue, not a circularity: the result is not equivalent by definition to its inputs, and no quoted reduction (Eq. X = Eq. Y, fitted parameter renamed as prediction, or self-citation chain) can be exhibited. Accordingly, the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The ledger is light on invented content: the paper introduces no new theoretical objects. Its load-bearing commitments are experimental: unreported hyperparameters (slot count, slot dimension, history length, GMM configuration) and three domain assumptions about protocol fairness, camera observability, and expert-trajectory quality. These assumptions, rather than any mathematical axiom, carry the central claim.

free parameters (4)
  • Slot count K = not reported
    Number of object slots in SAVi, a hand-chosen hyperparameter that sets decomposition granularity; never stated in Section 3.2, though it changes the policy input size.
  • Slot dimension D = not reported
    Embedding size of each slot in SAVi; not reported in Section 3.2.
  • History length H = not reported
    Number of past frames fed to the transformer observation trunk; chosen by hand and not reported in Section 3.2.
  • GMM head configuration = not reported
    Number of mixture components and training details of the Gaussian-mixture action head are unspecified in Section 3.2.
assumptions (4)
  • domain assumption In-domain pre-training of the representation is a fair protocol for comparing representation families.
    SAVi is pre-trained with a reconstruction objective on the same RoboShape data used for behavior cloning, while DINO and R3M are frozen off-the-shelf (Section 'Training' and Section 4.1). The conclusion assumes this asymmetry does not drive the results.
  • domain assumption A single front-facing camera view provides sufficient observability for the tasks.
    The environment uses one camera; the authors themselves attribute the total failure of Task 3 to single-view ambiguity (Section 4.2 and Figure 6), and the same limitation affects the general claim.
  • domain assumption Motion-planning expert trajectories are a valid and consistent demonstration distribution.
    2,000 expert trajectories per task are collected using oracle information and motion planning (Section 4.1); behavior cloning inherits whatever bias this distribution has.
  • standard math Slot-Attention and SAVi behave as described in the cited publications.
    The method relies directly on the published Slot-Attention mechanism [22] and SAVi [20] with no reformulation or independent check in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Is an object-centric representation beneficial for robotic manipulation ?." pith.science (2026). https://pith.science/paper/SEA2T5YA

@misc{pith2026250619408,
  author       = {Pith},
  title        = {Pith review of: Is an object-centric representation beneficial for robotic manipulation ?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SEA2T5YA}},
  note         = {Machine review of arXiv:2506.19408}
}
read the original abstract

Object-centric representation (OCR) has recently become a subject of interest in the computer vision community for learning a structured representation of images and videos. It has been several times presented as a potential way to improve data-efficiency and generalization capabilities to learn an agent on downstream tasks. However, most existing work only evaluates such models on scene decomposition, without any notion of reasoning over the learned representation. Robotic manipulation tasks generally involve multi-object environments with potential inter-object interaction. We thus argue that they are a very interesting playground to really evaluate the potential of existing object-centric work. To do so, we create several robotic manipulation tasks in simulated environments involving multiple objects (several distractors, the robot, etc.) and a high-level of randomization (object positions, colors, shapes, background, initial positions, etc.). We then evaluate one classical object-centric method across several generalization scenarios and compare its results against several state-of-the-art hollistic representations. Our results exhibit that existing methods are prone to failure in difficult scenarios involving complex scene structures, whereas object-centric methods help overcome these challenges.

Figures

Figures reproduced from arXiv: 2506.19408 by the authors.

Figure 1
Figure 1. Overview of the proposed framework. We introduce a new framework to evaluate different representations models for robot manipulation policy learning. Our main objective is to evaluate object-centric representation methods. To do so, the models can be trained following classical scheme through image reconstruction loss on in-domain data. Then, the model can be used as a frozen backbone for manipulation learning : it … view at source ↗
Figure 2
Figure 2. Example rollout of the tasks. Top row: Push cube to target; Middle row: Pick cube; Bottom row: Place cube in bin [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. L1 generalization - Unseen distractor colors [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: L2 generalization - Unseen background colors [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: L3 generalization - Unseen distractor sizes [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Failure case in the last task : the robot is blocked on top of the green box and do not realease the cube. Left : Input observation; Right : Model reconstruction Finally, the [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Overall success rate: mean success rate over the 3 tasks on the different generalization scenarios [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Object decomposition on rollout From left to right : input image, full reconstruction, slot decomposition 5 Conclusion We introduced a new simple framework which combines behavior cloning with object-centric representations. To validate the effectiveness of our approac…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 12 canonical work pages

  1. [1]

    Bengio, Y., Courville, A., Vincent, P.: Representation learning: A review and new perspectives (2014), https://arxiv.org/abs/1206.5538

    arXiv 2014

  2. [2]

    Brohan, A., et al.: Rt-1: Robotics transformer for real-world control at scale (2023), https://arxiv.org/abs/2212.06817

    arXiv 2023

  3. [3]

    Burgess, C.P., Matthey, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M., Lerchner, A.: Monet: Unsupervised scene decomposition and representation (2019), https://arxiv.org/abs/1901.11390

    arXiv 2019

  4. [4]

    Burns, K., Witzel, Z., Hamid, J.I., Yu, T., Finn, C., Hausman, K.: What makes pre- trained visual representations successful for robust manipulation? (2023),https: //arxiv.org/abs/2312.12444

    arXiv 2023

  5. [5]

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers (2021), https:// arxiv.org/abs/2104.14294

    arXiv 2021

  6. [6]

    Chi, C., Xu, Z., Feng, S., Cousineau, E., Du, Y., Burchfiel, B., Tedrake, R., Song, S.: Diffusion policy: Visuomotor policy learning via action diffusion (2024),https: //arxiv.org/abs/2303.04137

    arXiv 2024

  7. [7]

    Collaboration, E., O’Neill, A., et al.: Open x-embodiment: Robotic learning datasets and rt-x models (2024),https://arxiv.org/abs/2310.08864

    arXiv 2024

  8. [8]

    Dalal, M., Mandlekar, A., Garrett, C., Handa, A., Salakhutdinov, R., Fox, D.: Imitating task and motion planning with visuomotor transformers (2023)

    work page 2023

Show all 43 references
  1. [9]

    Elsayed, G.F., Mahendran, A., van Steenkiste, S., Greff, K., Mozer, M.C., Kipf, T.: Savi++: Towards end-to-end object-centric learning from real-world videos (2022), https://arxiv.org/abs/2206.07764

    2022 arXiv

  2. [10]

    https://doi.org/10.5281/ ZENODO.1207631, https://zenodo.org/record/1207631

    Ha, D., Schmidhuber, J.: World models (2018). https://doi.org/10.5281/ ZENODO.1207631, https://zenodo.org/record/1207631

    2018

  3. [11]

    Hafner, D., Lillicrap, T., Ba, J., Norouzi, M.: Dream to control: Learning behaviors by latent imagination (2020),https://arxiv.org/abs/1912.01603

    2020 arXiv

  4. [12]

    Chapin et al

    Haldar, S., Peng, Z., Pinto, L.: Baku: An efficient transformer for multi-task policy learning (2024), https://arxiv.org/abs/2406.07539 14 A. Chapin et al

    2024 arXiv

  5. [13]

    Haramati, D., Daniel, T., Tamar, A.: Entity-centric reinforcement learning for object manipulation from pixels (2024),https://arxiv.org/abs/2404.01220

    2024 arXiv

  6. [14]

    Heravi, N., Wahid, A., Lynch, C., Florence, P., Armstrong, T., Tompson, J., Ser- manet, P., Bohg, J., Dwibedi, D.: Visuomotor control in multi-object scenes using object-aware representations (2023),https://arxiv.org/abs/2205.06333

    2023 arXiv

  7. [15]

    In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=Sy2fzU9gl

    Higgins, I., Matthey, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., Mohamed, S., Lerchner, A.: beta-VAE: Learning basic visual concepts with a constrained variational framework. In: International Conference on Learning Representations (2017), https://openreview.net/foru...

    2017

  8. [16]

    Hu, Y., Wang, R., Li, L.E., Gao, Y.: For pre-trained vision models in motor control, not all policy learning methods are created equal (2023),https://arxiv.org/abs/ 2304.04591

    2023 arXiv

  9. [17]

    Jiang, J., Deng, F., Singh, G., Ahn, S.: Object-centric slot diffusion (2023),https: //arxiv.org/abs/2303.10834

    2023 arXiv

  10. [18]

    Kabra, R., Zoran, D., Erdogan, G., Matthey, L., Creswell, A., Botvinick, M., Lerch- ner, A., Burgess, C.P.: Simone: View-invariant, temporally-abstracted object rep- resentations via unsupervised video decomposition (2021), https://arxiv.org/ abs/2106.03849

    2021 arXiv

  11. [19]

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2022), https:// arxiv.org/abs/1312.6114

    2022 arXiv

  12. [20]

    Kipf, T., Elsayed, G.F., Mahendran, A., Stone, A., Sabour, S., Heigold, G., Jon- schkowski, R., Dosovitskiy, A., Greff, K.: Conditional object-centric learning from video (2022), https://arxiv.org/abs/2111.12594

    2022 arXiv

  13. [21]

    Kroemer, O., Niekum, S., Konidaris, G.: A review of robot learning for manip- ulation: Challenges, representations, and algorithms (2020),https://arxiv.org/ abs/1907.03146

    2020 arXiv

  14. [22]

    Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot atten- tion (2020), https://arxiv.org/abs/2006.15055

    2020 arXiv

  15. [23]

    Ma, Y.J., Sodhani, S., Jayaraman, D., Bastani, O., Kumar, V., Zhang, A.: Vip: Towards universal visual reward and representation via value-implicit pre-training (2023), https://arxiv.org/abs/2210.00030

    2023 arXiv

  16. [24]

    Majumdar, A., Yadav, K., Arnaud, S., Ma, Y.J., Chen, C., Silwal, S., Jain, A., Berges, V.P., Abbeel, P., Malik, J., Batra, D., Lin, Y., Maksymets, O., Rajeswaran, A., Meier, F.: Where are we in the search for an artificial visual cortex for embodied intelligence? (2024), https...

    2024 arXiv

  17. [25]

    Mandlekar, A., Xu, D., Wong, J., Nasiriany, S., Wang, C., Kulkarni, R., Fei-Fei, L., Savarese, S., Zhu, Y., Martín-Martín, R.: What matters in learning from offline human demonstrations for robot manipulation (2021),https://arxiv.org/abs/ 2108.03298

    2021 arXiv

  18. [26]

    Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal vi- sual representation for robot manipulation (2022),https://arxiv.org/abs/2203. 12601

    2022

  19. [27]

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W.,Howes,R.,Huang,P.Y.,Li,S.W.,Misra,I.,Rabbat,M.,Sharma,V.,Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P....

    2024 arXiv

  20. [28]

    arXiv preprint arXiv:2405.15916 (2024)

    Qian, J., Panagopoulos, A., Jayaraman, D.: Recasting generic pretrained vision transformers as object-centric scene encoders for manipulation policies. arXiv preprint arXiv:2405.15916 (2024)

    2024 arXiv

  21. [29]

    Radosavovic, I., Xiao, T., James, S., Abbeel, P., Malik, J., Darrell, T.: Real-world robot learning with masked visual pre-training (2022),https://arxiv.org/abs/ 2210.03109

    2022 arXiv

  22. [30]

    Seitzer, M., Horn, M., Zadaianchuk, A., Zietlow, D., Xiao, T., Simon-Gabriel, C.J., He, T., Zhang, Z., Schölkopf, B., Brox, T., Locatello, F.: Bridging the gap to real- world object-centric learning (2023),https://arxiv.org/abs/2209.14860

    2023 arXiv

  23. [31]

    Shang, J., Schmeckpeper, K., May, B.B., Minniti, M.V., Kelestemur, T., Watkins, D.,Herlant,L.:Theia:Distillingdiversevisionfoundationmodelsforrobotlearning (2024), https://arxiv.org/abs/2407.20179

    2024 arXiv

  24. [32]

    Singh, G., Deng, F., Ahn, S.: Illiterate dall-e learns to compose (2022),https: //arxiv.org/abs/2110.11405

    2022 arXiv

  25. [33]

    Singh, G., Wu, Y.F., Ahn, S.: Simple unsupervised object-centric learning for com- plex and naturalistic videos (2022),https://arxiv.org/abs/2205.14065

    2022 arXiv

  26. [34]

    org/abs/2410.00425

    Tao, S., Xiang, F., Shukla, A., Qin, Y., Hinrichsen, X., Yuan, X., Bao, C., Lin, X., Liu, Y., kai Chan, T., Gao, Y., Li, X., Mu, T., Xiao, N., Gurha, A., Huang, Z., Calandra, R., Chen, R., Luo, S., Su, H.: Maniskill3: Gpu parallelized robotics simulation and rendering for gene...

    2024 arXiv

  27. [35]

    Team, O.M., Ghosh, D., Walke, H., Pertsch, K., Black, K., Mees, O., Dasari, S., Hejna, J., Kreiman, T., Xu, C., Luo, J., Tan, Y.L., Chen, L.Y., Sanketi, P., Vuong, Q., Xiao, T., Sadigh, D., Finn, C., Levine, S.: Octo: An open-source generalist robot policy (2024),https://arxiv...

    2024 arXiv

  28. [36]

    Watters, N., Matthey, L., Bosnjak, M., Burgess, C.P., Lerchner, A.: Cobra: Data- efficient model-based rl through unsupervised object discovery and curiosity-driven exploration (2019), https://arxiv.org/abs/1905.09275

    2019 arXiv

  29. [37]

    Wu, Z., Dvornik, N., Greff, K., Kipf, T., Garg, A.: Slotformer: Unsupervised visual dynamics simulation with object-centric models (2023),https://arxiv.org/abs/ 2210.05861

    2023 arXiv

  30. [38]

    Wu, Z., Hu, J., Lu, W., Gilitschenski, I., Garg, A.: Slotdiffusion: Object-centric generative modeling with diffusion models (2023),https://arxiv.org/abs/2305. 11281

    2023

  31. [39]

    In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)

    Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)

    2020

  32. [40]

    Yoon, J., Wu, Y.F., Bae, H., Ahn, S.: An investigation into pre-training object- centric representations for reinforcement learning (2023), https://arxiv.org/ abs/2302.04419

    2023 arXiv

  33. [41]

    Zadaianchuk, A., Seitzer, M., Martius, G.: Object-centric learning for real-world videosbypredictingtemporalfeaturesimilarities(2023), https://arxiv.org/abs/ 2306.04829

    2023 arXiv

  34. [42]

    Zhang, C., Gupta, A., Zisserman, A.: Is an object-centric video representation beneficial for transfer? (2022),https://arxiv.org/abs/2207.10075

    2022 arXiv

  35. [43]

    Zhu, Y., Joshi, A., Stone, P., Zhu, Y.: Viola: Imitation learning for vision-based manipulation with object proposal priors (2023),https://arxiv.org/abs/2210. 11339

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.