{"id":"0bb000b6-4680-466c-b2bc-9fe4f5d546c0","arxiv_id":"2505.11563","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Slot-based object-centric representations, especially a video model pretrained on robot data, improve policy generalization under visual distribution shifts in simulated and real-world manipulation tasks.","lead":"This paper compares object-centric, global, and dense visual encoders for robot manipulation policies, and reports that slot-based representations generalize better under lighting, texture, and distractor changes. The strongest gains come from a video-slot model retrained on a mixture of robot demonstration datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gains attributed to object-centric slots are confounded with robot-data pretraining and an unreported DINOv2 dense baseline; the OCR-specific claim is not yet isolated.","rationale":"The reader's weakest assumption already identifies the central confound: VIDEOSAUR* differs from the dense baselines in backbone, slot-attention mechanism, temporal transformer, and robot-data pretraining, with no ablation isolating slot attention. My stress-test agrees and sharpens the concern. The single most load-bearing issue is attribution: the paper's headline claim requires that object-centric structure, not the concurrent changes in pretraining data and architecture, causes the generalization improvement. This is load-bearing because the entire contribution depends on it; if a dense DINOv2 baseline or a non-slot control performs equally well, the paper reduces to an engineering result about pretraining on robot video, not a finding about object-centric representations. The evidence currently in the paper cannot settle this, because DINOv2 dense numbers were collected but not reported (Appendix E), and Table 6 shows pretraining data alone accounts for large gains. I do not think this warrants rejection: the benchmark is useful, the real-world comparisons are suggestive, and the missing baselines are feasible. The existing CONDITIONAL verdict is appropriate, so no verdict change is needed. The concrete test I propose would directly settle the attribution question: add the omitted dense DINOv2 baseline and a slot-attention ablation, both under identical policy training and evaluation protocols.","tokens_in":17010,"tokens_out":3935,"duration_ms":42542,"concrete_test":"Re-run the LIBERO-90 and real-world comparisons with a dense DINOv2 ViT-B14 baseline: feed patch tokens into the same BAKU/ACT observation trunk with identical training and rollout counts, and report success rates with confidence intervals. Add a control that keeps VIDEOSAUR*'s architecture but replaces Slot Attention with K learned linear projections of the DINOv2 patch features (no iterative competition), so the slot mechanism is isolated from the backbone and pretraining. If dense DINOv2 or the linear-projection control matches VIDEOSAUR* within sampling error, the OCR-specific claim is unsupported; if VIDEOSAUR* remains clearly ahead, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central inference is that the object-centric inductive bias causes the observed generalization gains. But VIDEOSAUR* differs from every dense baseline along multiple axes simultaneously. First, its slot-attention module was pretrained on 188k robot trajectories from BridgeData V2, Fractal, and DROID (Section 4, 'Robotic pre-training'); Table 6 shows this robot-mixture pretraining alone lifts VIDEOSAUR from 0.58 to 0.70 in real-world and from 0.77 to 0.86 in LIBERO, so a large part of the headline gain may be pretraining-data alignment rather than object-centric structure. Second, VIDEOSAUR* uses a DINOv2 ViT-B14 backbone, but Appendix E states that DINOv2 was only reported in its global form because 'the Global representation was always outperforming the other alternative'; the dense DINOv2 result is never shown. Without the same-backbone dense baseline, one cannot tell whether slot attention adds anything over plain DINOv2 patch features. Third, VIDEOSAUR* adds a temporal transformer, and Section 5.1 credits temporal dynamics with +9 and +26 points over DINOSAUR* in LIBERO and LeRobot; those gains are not object-centric per se. The abstract's 'even without task-specific pretraining' is also strained: the robot mixture is manipulation-specific pretraining, and the non-robot-pretrained OCR variants do not consistently beat Theia (MetaWorld Table 2: DINOSAUR 0.46, VIDEOSAUR 0.41 vs Theia 0.47). Thus the load-bearing premise, that OCR structure rather than pretraining, backbone, or temporal architecture drives the result, remains untested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether object-centric representations (OCRs), specifically slot-based encoders DINOSAUR and VIDEOSAUR, improve the learning and generalization of visuomotor policies compared with global and dense visual representations. The authors introduce robot-pretrained variants DINOSAUR* and VIDEOSAUR* obtained by training the slot-attention module on a mixture of BridgeData V2, Fractal, and DROID, and evaluate all models on MetaWorld, LIBERO-90, and five real-world LeRobot tasks under in-domain and shifted conditions (distractors, textures, lighting). The central claim is that OCR-based policies outperform dense and global representations in generalization settings, even without task-specific pretraining, with VIDEOSAUR* reported as the strongest method.","tokens_in":17383,"tokens_out":2409,"duration_ms":24543,"significance":"If the central claim is established, the paper would provide a useful benchmark and a concrete argument for object-centric inductive biases in robotic manipulation, supported by a unified framework, frozen-encoder comparisons across simulation and real hardware, and open-source release of the evaluation code. The paper also contributes a practical pretraining recipe for slot-attention models on large robot datasets and reports task-level performance. However, the evidence as presented does not yet isolate the OCR inductive bias from several confounds, and some headline numbers rest on very small real-world evaluation budgets.","major_comments":[{"comment":"The headline comparison between VIDEOSAUR* and the dense baselines is confounded by the robot-mixture pretraining. Table 6 shows that pretraining on the robot mixture improves VIDEOSAUR from 0.77 to 0.86 in LIBERO and from 0.58 to 0.70 in the real-world setup, gains of the same magnitude as the reported OCR advantage over dense baselines. Without a dense or global encoder trained on the same robot-mixture data, the observed gains cannot be attributed specifically to object-centric structure rather than to pretraining-data alignment. Please add an ablation that trains a non-OCR baseline on the same robot mixture, or otherwise remove the attribution of these gains to the OCR inductive bias.","section":"Section 4, 'Robotic pre-training'; Table 6"},{"comment":"The DINOv2 dense representation is mentioned but never reported: the appendix states that 'the Global representation was always outperforming the other alternative,' yet no dense DINOv2 numbers are shown anywhere. Since VIDEOSAUR* uses a DINOv2 ViT-B14 backbone with slot attention on top, the missing DINOv2 dense baseline is exactly the control needed to determine whether slot attention adds anything over the frozen DINOv2 patch features. Please report the DINOv2 dense result in all tables, or justify its omission with explicit numbers.","section":"Appendix E, 'Baselines details'"},{"comment":"The real-world results are based on only 10 rollouts per task and are reported without error bars or per-seed variation, despite the stated protocol of three random seeds. Table 3 shows differences such as VIDEOSAUR* 0.44 overall versus VIDEOSAUR 0.40 which are likely within the noise of 10 rollouts per condition, and Figure 3 has no error bars at all. The claim that OCRs 'consistently' generalize better in the real world needs confidence intervals, more rollouts, or per-seed results; otherwise the 70% versus 50% headline comparison is not robustly supported.","section":"Section 5.1; Table 3; Figure 3"},{"comment":"The claim that OCRs outperform dense and global representations 'even without task-specific pretraining' is not supported by the presented comparisons. The robot-mixture pretraining is manipulation-specific pretraining, so the starred models do not satisfy the 'without task-specific pretraining' condition. Moreover, in Table 2 the non-robot-pretrained OCR models, DINOSAUR at 0.46 and VIDEOSAUR at 0.41, do not beat Theia at 0.47 on MetaWorld overall. Please either qualify the claim to distinguish robot-pretrained and non-robot-pretrained OCR variants, or provide evidence that non-robot-pretrained OCRs consistently surpass strong dense baselines.","section":"Abstract; Section 5.1; Table 2"}],"minor_comments":[{"comment":"The dimensions in Eq. (1) are inconsistent: the text defines Q in R^{NxD} but writes K in R^{KxD}, while Slot Attention normally projects queries from K slots and keys from N features; please clarify the notation so that the softmax dimensions match the description.","section":"Section 3, Eq. (1)"},{"comment":"The captions of Figure 7 and Figure 8 say '12 slots' and '8 slots', respectively, while Table 4 specifies 10 slots for both DINOSAUR and VIDEOSAUR; please reconcile these numbers.","section":"Appendix G and Appendix H"},{"comment":"There are several typos in this section, including 'Convolutionnal' and 'alse' for 'also'; a proofreading pass would improve readability.","section":"Section 3, 'Object-centric representation for videos'"},{"comment":"In Table 8, the VIDEOSAUR row is cited as '[18]' but should reference [41]; please correct the citation.","section":"Table 8"},{"comment":"The sentence 'In real-world evaluations as can be seen in Table 3' refers to the table that follows, but the preceding sentence also references Table 2; please make the table references unambiguous.","section":"Section 5.2, paragraph 2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and relevant question for the robotics community, and the proposed benchmark has clear value. The main risk is that the reported gains are currently attributable to a combination of robot-data pretraining, temporal modeling, and backbone choice rather than to the object-centric inductive bias. The requested ablations and uncertainty reporting are feasible within the scope of a revision, so I see this as a major-revision rather than a reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a useful benchmark paper with a headline claim that outruns its evidence. The benchmark itself — comparing global, dense, and object-centric encoders across MetaWorld, LIBERO-90, and a small real-world SO-100 suite — is a solid addition to the visuomotor-policy literature. The robot-mixture pretraining of slot attention (188k trajectories from Bridge, Fractal, DROID) is a genuinely new piece and clearly helps: Table 6 shows it lifts VIDEOSAUR from 0.58 to 0.70 in real-world and 0.77 to 0.86 in LIBERO.\n\nThe soft spot is attribution. The paper claims OCRs improve generalization \"even without task-specific pretraining,\" but the evidence doesn't isolate the object-centric inductive bias. VIDEOSAUR* differs from the dense baselines by backbone (DINOv2), slot attention, a temporal transformer, and robot pretraining of the slot module. The dense DINOv2 result is withheld (Appendix E says global always outperformed), so we can't separate slot attention from plain DINOv2 patch features. The non-robot-pretrained OCR variants don't consistently beat Theia (MetaWorld Table 2: DINOSAUR 0.46, VIDEOSAUR 0.41 vs Theia 0.47). And the real-world numbers are 10 rollouts per task without error bars; even the stated 3-seed protocol isn't honored in Figure 3 or Table 6.\n\nNone of this kills the paper. The benchmark is useful, the pretraining data ablation is informative, and the limitations section is honest. But the central claim needs an ablation that holds backbone and pretraining fixed and turns slot attention on/off, plus the dense DINOv2 baseline and error bars (or at least confidence intervals on the real-world numbers). The \"without task-specific pretraining\" phrasing should be dropped or heavily qualified, since the robot mixture is manipulation-specific pretraining.\n\nWho's this for? Anyone working on visual front-ends for visuomotor policies will find the comparison worth reading. It deserves peer review — the empirical scope is beyond a desk reject — but it should come back with major revisions. If the authors add the missing controls, it could be a useful reference point.\n\nRecommendation: send to review, but expect the confound issue to dominate the referee reports. Cite if you work on this area, but don't cite the causal claim — cite the benchmark.","headline":"Useful benchmark, but the headline causal claim about object-centric structure is confounded with robot-data pretraining and an unreported DINOv2 dense baseline.","tokens_in":17917,"tokens_out":2235,"would_cite":true,"duration_ms":20126,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Object-centric vision lifts robot manipulation generalization","keywords":["object-centric representations","slot attention","robot manipulation","policy generalization","imitation learning","visual representation learning","distribution shift","VIDEOSAUR"],"falsifier":"Run the VIDEOSAUR* pipeline with the slot-attention module replaced by a mean-pooling or learned pooling of the same frozen DINOv2 features, keeping the temporal transformer and robot-mixture pretraining; if that pooled variant still reaches roughly 70% real-world success and similar LIBERO numbers, the paper's attribution of the gains to object-centric structure is falsified.","tokens_in":16853,"feed_emoji":"🤖","tokens_out":5294,"duration_ms":48610,"temperature":0.7,"pith_summary":"The paper sets out to test whether object-centric representations—encodings that split an image into a small set of entity slots rather than a global vector or dense patch map—give robot manipulation policies better generalization under visual shifts such as new lighting, textures, and distractors. Across two simulation benchmarks and a five-task real-world setup, the authors find that slot-based encoders, especially their robot-pretrained VIDEOSAUR variant, outperform dense and global baselines in out-of-distribution conditions and match or beat them in-domain. The headline result is a 70% real-world success rate for VIDEOSAUR* versus 50% for the best dense baseline. The authors argue this makes object-centric encoders a promising direction for designing robust visuomotor policies in real-world environments.","feed_headline":"Object-centric vision lifts robot manipulation generalization","feed_subtitle":"Slot-based encoders beat dense and global features under new lighting, textures, and distractors.","key_machinery":"The load-bearing mechanism is Slot Attention with a frozen DINOv2 vision backbone, as realized in DINOSAUR and extended in VIDEOSAUR. Slot Attention is an iterative cross-attention module that compresses N dense patch features into K slot vectors through a softmax renormalization over slots, so each slot specializes on one entity. VIDEOSAUR adds a transformer predictor that initializes slots at time t from slots at t−1 plus a temporal consistency loss, and the paper's VIDEOSAUR* retrains the slot-attention module on a mixture of robot manipulation videos to align the slots with manipulation dynamics. These slot vectors are fed, frozen, into transformer-based policy heads (BAKU in simulation, ACT in the real world), replacing the usual global or dense feature input.","core_discovery":"The central claim is that OCR-based policies outperform dense and global representations in generalization settings, even without task-specific pretraining. In the paper's comparison, VIDEOSAUR*—a slot-attention video model with a DINOv2 backbone, a temporal transformer predictor, and slot attention retrained on a mixture of 188k robot trajectories—achieves the highest average success in LIBERO-90 and in the real-world suite, reaching 70% success versus 50% for the best dense baseline, while remaining competitive in MetaWorld. The paper attributes this to the inductive bias of slot attention: decomposing the scene into discrete entities lets the policy ignore task-irrelevant background and stay robust to appearance changes. The authors also report that robot-data pretraining and temporal dynamics modeling each contribute large gains, with VIDEOSAUR* beating DINOSAUR* by 9 and 26 points in LIBERO and the real-world suite, respectively.","pith_inferences":["A matched ablation that replaces the slot-attention module with pooling of the same frozen DINOv2 features, keeping the temporal transformer and robot-mixture pretraining, would settle whether the object-centric structure itself or the extra components cause the gains.","The 'without task-specific pretraining' phrasing is definitionally fragile: all three pretraining sources are manipulation datasets, so a stricter reading is that the gains survive when the objective is reconstruction rather than action prediction.","The paper's own slot visualizations suggest that grounding slots semantically, for example with affordances, could reduce distractor capture and is a natural next test for the approach.","If object-centric encoders are adopted more widely, the practical benchmark to watch is whether they continue to dominate when dense baselines are given the same backbone, data, and temporal modeling."],"forward_implications":["If the central claim holds, robot policy designers can expect slot-based encoders to provide more robust performance under lighting, texture, and distractor shifts than dense or global encoders.","Pretraining the slot-attention module on large robot video collections is a key lever: VIDEOSAUR* adds 10–13 points in mean success over VIDEOSAUR across the three environments.","Temporal dynamics in the object-centric encoder matter: VIDEOSAUR* beats DINOSAUR* by 9 points in LIBERO and 26 points in the real-world suite when both use the same robot-mixture pretraining.","Object-centric models remain competitive in-domain while winning out-of-distribution, so switching to them does not sacrifice standard performance in these benchmarks."],"supporting_citations":[{"why":"Defines Slot Attention, the core iterative cross-attention mechanism that produces the object-centric slot vectors.","marker":"[17]"},{"why":"Introduces DINOSAUR, the real-world object-centric model that reconstructs DINO features and provides the base OCR architecture used here.","marker":"[18]"},{"why":"Introduces VIDEOSAUR, the video extension with a temporal transformer predictor and feature-similarity loss that the paper adapts and robot-pretrains.","marker":"[41]"},{"why":"Supplies the frozen DINOv2 backbone used by the object-centric encoders.","marker":"[27]"},{"why":"Provides the BAKU policy architecture used in the simulation experiments.","marker":"[1]"},{"why":"Provides the ACT policy architecture modified for the real-world experiments.","marker":"[44]"},{"why":"Motivates the robot-mixture pretraining by showing dataset diversity and quality matter for visuomotor pretraining.","marker":"[30]"},{"why":"One of the three robot datasets (BridgeData V2) used to retrain the slot-attention module.","marker":"[50]"},{"why":"One of the three robot datasets (Fractal) used to retrain the slot-attention module.","marker":"[2]"},{"why":"One of the three robot datasets (DROID) used to retrain the slot-attention module.","marker":"[51]"}],"fun_headline_variants":["Slot attention beats dense features for robot generalization","Object-centric slots improve robot policy generalization","Robot policies gain from object-centric scene decomposition","Object-centric encoders beat global features for robot policies","Slot-attention policies generalize better in robot tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the performance gap comes from the object-centric inductive bias itself, but the best model differs from the dense baselines in several simultaneous ways—a DINOv2 backbone, a temporal transformer, and slot-attention pretraining on 188k robot trajectories—so no single factor is isolated, and calling those robot datasets 'not task-specific' is itself an assumption.","fun_headline_variants_meta":{"raw":{"variants":["Slot attention beats dense features for robot generalization","Object-centric slots improve robot policy generalization","Robot policies gain from object-centric scene decomposition","Object-centric encoders beat global features for robot policies","Slot-attention policies generalize better in robot tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2862,"prompt_tokens":879,"completion_tokens":1983,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":1915}},"tokens_in":495,"tokens_out":1983,"duration_ms":13398,"temperature":1.0,"reasoning_tokens":1915,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:00:22.875571+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the VIDEOSAUR* pipeline with the slot-attention module replaced by a mean-pooling or learned pooling of the same frozen DINOv2 features, keeping the temporal transformer and robot-mixture pretraining; if that pooled variant still reaches roughly 70% real-world success and similar LIBERO numbers, the paper's attribution of the gains to object-centric structure is falsified.","supporting_citations":[{"cited_title":"An Unbiased Look at Datasets for Visuo-Motor Pre-Training","cited_arxiv_id":"2310.09289","evidence_quote":"Motivates the robot-mixture pretraining by showing dataset diversity and quality matter for visuomotor pretraining."}],"review_version":1}