{"id":"33613909-eac3-4645-8ab7-810db0ea1ddb","arxiv_id":"2606.21216","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Large-scale real-world study demonstrates that bottlenecking pre-trained ViT encoders to a scalar per patch produces interpretable affordance features useful for navigation policies.","lead":"This paper reports results from 966 real-world robot navigation episodes covering 24 km to test pre-trained visual encoders. A smart generalist might read it to learn which vision components matter most for practical robot deployment in buildings.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Single-building evaluation (one environment, 966 episodes) undercuts claims that bottlenecking produces generally useful affordance-linked features.","rationale":"The reader’s weakest_assumption directly identifies the same generalization risk that is load-bearing for the strongest_claim. Because the full text confirms the single-site data collection, the concern stands and the UNVERDICTED verdict with low confidence remains appropriate.","tokens_in":1702,"tokens_out":281,"duration_ms":11833,"concrete_test":"Re-run the identical bottlenecking procedure and feature-visualization analysis on a second, structurally different building (or on a held-out floor with different corridor widths and obstacle distributions); measure whether the same affordance-linked interpretations appear and whether policy performance transfers without retraining the encoder.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the interpretable features emerging from principled spatial bottlenecking are linked to affordances in a way that supports real-world navigation. All 966 episodes and the associated distillation/bottlenecking experiments occur inside a single building. No cross-building or cross-environment transfer results are reported for the feature interpretability or affordance linkage. If the observed features are artifacts of that building’s geometry, lighting, or layout rather than general affordance signals, the emergence claim does not hold outside the tested site.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper reports a large-scale real-world evaluation of 966 static point-goal navigation episodes (24 km total) inside a single building. It compares state-of-the-art pre-trained visual encoders, heterogeneous multi-teacher distillation, and a principled spatial bottlenecking procedure; the central claim is that the bottlenecking step produces interpretable features linked to affordances. Additional claims are that RGB-only policy training is suboptimal (demonstrated via privileged-information fine-tuning) and that the study clarifies which computer-vision components matter for real-world navigation.","tokens_in":1794,"tokens_out":473,"duration_ms":15698,"significance":"If the emergence of affordance-linked features from spatial bottlenecking generalizes, the result would supply concrete guidance for encoder design in robotics. The scale of the real-world data collection is a clear strength; the work also supplies an empirical comparison of distillation strategies that is rarely performed at this volume outside simulation.","major_comments":[{"comment":"Abstract and § on experimental results: the central claim that bottlenecking yields 'interpretable features linked to affordances' is supported only by experiments inside one building; no cross-building or cross-environment transfer results are reported for either feature interpretability or navigation performance, which is load-bearing for the generalization implied by the title and abstract.","section":"Abstract and experimental evaluation"},{"comment":"Abstract: quantitative support for all reported outcomes (success rates, ablation deltas, feature-interpretability metrics) is absent; the manuscript therefore provides no error bars, confidence intervals, or statistical tests that would allow a reader to assess whether the observed effects exceed environment-specific noise.","section":"Abstract"}],"minor_comments":[{"comment":"The description of the spatial bottlenecking procedure would benefit from an explicit equation or diagram showing how the per-patch scalar is computed and how it is inserted into the ViT forward pass.","section":null},{"comment":"The manuscript should clarify the precise definition of 'heterogeneous multi-teacher distillation' (number of teachers, loss weighting, and which layers are distilled) so that the ablation can be reproduced.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive feedback and for recognizing the scale of the real-world evaluation. We address each major comment below, clarifying the scope of our claims and committing to revisions where appropriate.","responses":[{"response":"The study is explicitly conducted within a single building, as stated in the abstract and experimental section. The title's reference to 'the real world' contrasts with simulation rather than claiming multi-environment generalization. Feature interpretability is evidenced through visualizations and affordance correlations observed in this environment. We will revise the abstract and discussion sections to explicitly delimit the single-building scope and remove any phrasing that could imply broader transfer.","revision_made":"partial","referee_comment":"[Abstract and experimental evaluation] Abstract and § on experimental results: the central claim that bottlenecking yields 'interpretable features linked to affordances' is supported only by experiments inside one building; no cross-building or cross-environment transfer results are reported for either feature interpretability or navigation performance, which is load-bearing for the generalization implied by the title and abstract."},{"response":"The current manuscript reports aggregate outcomes from 966 episodes but omits error bars and statistical tests. We will add these elements, including confidence intervals and significance tests for success rates and ablation deltas, in the revised version to quantify variability.","revision_made":"yes","referee_comment":"[Abstract] Abstract: quantitative support for all reported outcomes (success rates, ablation deltas, feature-interpretability metrics) is absent; the manuscript therefore provides no error bars, confidence intervals, or statistical tests that would allow a reader to assess whether the observed effects exceed environment-specific noise."}],"tokens_in":1362,"tokens_out":363,"duration_ms":13374,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's clearest contribution is the volume of real robot data: 966 point-goal navigation episodes totaling 24 km inside one building. They use that to compare several pre-trained ViTs, test heterogeneous multi-teacher distillation, and apply a spatial bottleneck that reduces each patch to a scalar. The distillation and bottleneck steps are extensions of known ideas, but running them at this physical scale is uncommon and gives concrete evidence on what survives when you move from sim to actual hallways and lighting.\n\nThe work does well on the practical side. They also show that policies fine-tuned from privileged information outperform those trained on RGB alone, which aligns with what most people expect but is still worth documenting with real runs. The spatial bottlenecking producing interpretable outputs is the part that could matter for encoder design if it holds up.\n\nThe main limitation is exactly what the stress-test note flags: every episode stays in the same building. Without any cross-site transfer results, the claim that the bottlenecked features are linked to general affordances rather than local geometry or illumination is hard to assess. The abstract gives no numbers, error bars, or details on how interpretability was measured, so the emergence story rests on qualitative observation in one environment. That makes the generalization step the weakest part.\n\nThis is worth a serious referee for groups working on visual encoders for navigation. The real-world mileage is rare enough that reviewers can help tighten the claims around what transfers and what the bottleneck actually captures. I would bring it to a reading group to talk through the evaluation design, but I would not cite the affordance-emergence result until we see data from at least one more site.","headline":"Real-world scale from 966 episodes is the real asset here, but the single-building setup limits what we can take away about general affordance features from bottlenecking.","tokens_in":2244,"tokens_out":412,"would_cite":false,"duration_ms":15095,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Bottlenecking pre-trained ViTs to one scalar per patch yields features for real-world robot navigation.","keywords":["visual encoders","vision transformers","robot navigation","bottlenecking","affordances","multi-teacher distillation","real-world robotics","point goal navigation"],"falsifier":"Navigation success rate drops sharply when the same scalar-per-patch policies are deployed in a second building with different layout or lighting while full-resolution encoders maintain performance.","tokens_in":2609,"feed_emoji":"🤖","tokens_out":548,"duration_ms":16604,"temperature":0.7,"pith_summary":"The paper tests state-of-the-art visual encoders on 966 real-world navigation episodes totaling 24 km and finds that a spatially structured reduction of ViT outputs to one scalar per patch is sufficient for training effective policies. Heterogeneous multi-teacher distillation first equips the encoders with complementary skills, after which the bottleneck causes interpretable features tied to environmental affordances to emerge. Policies using only RGB inputs without this reduction prove suboptimal, as finetuning on privileged information improves performance.","feed_headline":"One scalar per ViT patch enables real robot navigation","feed_subtitle":"Bottleneck after distillation produces affordance-linked features across 24 km of episodes in a building.","key_machinery":"The scalar-per-patch bottleneck applied to pre-trained ViT encoders after multi-teacher distillation.","core_discovery":"Reducing pre-trained ViT encoders to a single scalar per image patch through principled bottlenecking, after heterogeneous distillation, produces navigation policies that succeed on static point-goal tasks in a real building while generating features that can be interpreted as affordances.","pith_inferences":["The same per-patch scalar reduction could be tested on other robotics tasks that rely on pre-trained vision backbones.","Measuring how the scalar features correlate with specific actions across varied buildings would test the affordance interpretation.","The approach suggests that explicit spatial structure in the bottleneck matters more than raw feature dimensionality for policy learning."],"forward_implications":["Policies trained on the reduced scalar features complete long-distance navigation episodes in real buildings.","Distillation from multiple heterogeneous teachers equips encoders with complementary visual skills.","Finetuning shows that policies pre-trained on privileged information outperform those trained on RGB alone.","The bottleneck step causes features linked to affordances to appear without explicit supervision."],"fun_headline_variants":["Scalar per ViT patch enables real robot navigation","ViT scalars per patch enable building navigation","Per-patch scalar from ViTs aids real navigation","Patch scalars after distillation enable robot navigation"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Results from 966 episodes inside one building will generalize to broader real-world navigation conditions.","fun_headline_variants_meta":{"raw":{"variants":["Scalar per ViT patch enables real robot navigation","ViT scalars per patch enable building navigation","Per-patch scalar from ViTs aids real navigation","Patch scalars after distillation enable robot navigation"]},"model":"grok-4.3","cost_usd":0.006012,"raw_usage":{"total_tokens":2835,"prompt_tokens":645,"num_sources_used":0,"completion_tokens":54,"cost_in_usd_ticks":60124500,"prompt_tokens_details":{"text_tokens":645,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2136,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":645,"tokens_out":54,"duration_ms":14340,"temperature":1.0,"reasoning_tokens":2136,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T14:17:50.121196+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Navigation success rate drops sharply when the same scalar-per-patch policies are deployed in a second building with different layout or lighting while full-resolution encoders maintain performance.","supporting_citations":[],"review_version":1}