{"id":"6e48c0c9-0b79-438d-8c83-cf4b1516a135","arxiv_id":"2608.13448","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"EDD splits scene images into environmental and social-agent views, trains a dual-branch network with replay, and improves prediction of socially appropriate robot actions across six indoor domains in a domain-incremental continual learning setting.","lead":"This paper introduces EDD, a framework that separates the environment and the people in a scene so a robot can keep learning socially appropriate actions as it encounters new rooms. The method combines panoptic segmentation with two neural-network branches and replay memory, and reports improvements over continual learning baselines on synthetic home and office scenes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed benefit of explicit environment/social disentanglement rests on sub-noise differences in Table VII, with no significance test; the central RQ1 claim is not yet empirically supported.","rationale":"The reader's weakest assumption concerns the validity of Eq. (2) and whether masking destroys task-relevant information. That is a genuine theoretical risk, and the paper explicitly acknowledges in Future Work that robustness to imperfect panoptic segmentation is untested. However, the more immediately load-bearing problem is empirical: the comparison that isolates the contribution of explicit decomposition—EDD vs NoDec and vs single-branch CL—shows differences smaller than the reported cross-validation standard deviations, with no significance testing anywhere in the paper. The reader's own verdict flags the small advantage and lack of significance testing, so I partially agree, but I weight the empirical non-significance more heavily than the Eq. (2) decomposition-risk concern because the paper presents the RQ1 conclusion as settled despite the table-level evidence. The recommended CONDITIONAL verdict remains appropriate: the paper is reproducible, well-structured, and the framework is plausible, but the central claim needs statistical support and a direct test of whether the small observed gains are real. A paired significance test is a cheap, concrete check that would settle the matter.","tokens_in":15925,"tokens_out":1515,"duration_ms":15956,"concrete_test":"Run a paired significance test, e.g., a paired bootstrap or Wilcoxon signed-rank test over the 5 folds (or per-domain paired deltas over the 30 domain-fold pairs), comparing EDD Bbox vs NoDec and vs single-branch CL on RMSE, PCC, and CCC. Report the per-fold EDD Bbox minus NoDec deltas with a 95% confidence interval. If the CI for the RMSE delta includes 0, the claim that explicit decomposition improves CL is not supported by the current evidence.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that explicit environmental-social disentanglement improves domain-incremental CL. The load-bearing comparison is EDD (Bbox) vs. the NoDec control in Table VII: overall RMSE 0.78 vs 0.79, PCC 0.55 vs 0.53, CCC 0.38 vs 0.36. Given reported SDs of 0.03, 0.05, and 0.04, every advantage is under one standard deviation of 5-fold CV noise. No significance test, confidence interval, or per-fold breakdown is provided. On pooled Office domains the gap is 0.76 vs 0.77 RMSE and 0.56 vs 0.55 PCC; on Home it is 0.88 vs 0.89 RMSE and 0.53 vs 0.48 PCC. Table VI shows the single-branch CL control at 0.79 RMSE / 0.53 PCC / 0.36 CCC, nearly indistinguishable from EDD CL, while single-branch joint training achieves 0.75 RMSE / 0.58 PCC—better than EDD CL. Thus the evidence does not isolate explicit decomposition as the cause of EDD's advantage over baselines; the effect could stem from architecture, replay, or initialization. The paper's own RQ2 discussion concedes 'all methods perform relatively similarly,' yet the RQ1 conclusion asserts 'explicit environmental-social separation supports CL outcomes' without statistical support. The future-work acknowledgment that robustness to imperfect panoptic segmentation is untested further exposes the decomposition premise as unvalidated under the stated segmentation failure mode.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses domain-incremental continual learning (CL) for socially appropriate robot actions. It proposes the Explicit Disentanglement Dual-Branch (EDD) framework, which uses panoptic segmentation to separate a scene into environmental and social-agent views, processes them through two branches of a shared network, and trains with experience replay. The authors evaluate EDD on six indoor domains from the merged OfficeDB/MannersDB+ datasets, compare it against robotics CL baselines, domain-incremental methods, and vision-language models, and additionally study the effect of different disentanglement strategies and domain orderings. The central claim is that explicitly separating environmental and social cues improves domain-incremental CL of action appropriateness.","tokens_in":16245,"tokens_out":4262,"duration_ms":42823,"significance":"If the central claim is supported, the paper addresses a relevant and underexplored gap: domain-incremental continual learning for socially appropriate robot actions. The framework is clearly described, the code is publicly available, and the evaluation includes standard 5-fold cross-validation, multiple regression metrics, a forgetting metric, ablations, and a latency analysis. The comparison against several baselines, including zero-shot VLMs, is a useful contribution. The main weakness is that the claimed benefit of explicit disentanglement over a no-decomposition control is quantitatively very small and not supported by statistical tests, which undermines the paper's primary research conclusion. The manuscript's own discussion acknowledges that all decomposition strategies perform relatively similarly, creating an internal inconsistency with the RQ1 conclusion.","major_comments":[{"comment":"The evidence does not support the RQ1 conclusion that explicit environmental-social separation supports CL outcomes. The load-bearing comparison is EDD (Bbox) versus the NoDec control: overall RMSE 0.78 vs 0.79, PCC 0.55 vs 0.53, CCC 0.38 vs 0.36, with standard deviations of 0.03/0.02, 0.05/0.05, and 0.04/0.05 respectively. Every advantage is under one standard deviation of the 5-fold cross-validation noise, and no significance test, confidence interval, or per-fold breakdown is provided. Table VI further shows that the single-branch CL control achieves 0.79 RMSE, 0.53 PCC, and 0.36 CCC, nearly indistinguishable from EDD CL, while single-branch joint training achieves 0.75 RMSE and 0.58 PCC, which is better than EDD CL. The paper's own RQ2 discussion states that 'all methods perform relatively similarly,' yet the RQ1 conclusion asserts that explicit separation supports CL outcomes. Please add paired statistical tests across the five folds, report effect sizes or confidence intervals, and temper the central claim accordingly.","section":"Section V-B / Table VII"},{"comment":"The decomposition premise p(y|x) ≈ p(y|x_E, x_S) is load-bearing for the entire method, but it is not validated under imperfect panoptic segmentation. The paper's future-work section acknowledges that robustness to imperfect panoptic segmentation is untested. If the segmentation model misses agents or produces false detections, the environmental and social views constructed by the transformations T_E and T_S discard information the network needs. The manuscript provides no evaluation of segmentation accuracy on the six domains and no sensitivity analysis. Please report segmentation quality on the evaluation data or introduce controlled noise experiments (e.g., dropping or adding masks) to demonstrate that the decomposition is sufficiently lossless for the proposed benefit to be plausible.","section":"Section III-C / Eq. (2)"},{"comment":"The baseline comparison is weakened by the extremely poor performance of FedLGR (RMSE 2.091, near-zero PCC and CCC), which suggests that the adaptation of this federated continual-learning method to the proposed domain-incremental setup may be suboptimal. The paper states that all baselines shared identical training procedures, but it does not describe in sufficient detail how FedLGR's client-based formulation was extended to this setting or whether its hyperparameters were tuned comparably. Please provide the adaptation details and, if possible, a stronger or more fairly tuned robotics CL baseline; otherwise, the claim that EDD outperforms state-of-the-art baselines rests partly on an unfavorable baseline configuration.","section":"Section IV-B / Table V"}],"minor_comments":[{"comment":"The discussion sentence 'all methods perform relatively similarly' (Section V-B) conflicts with the later conclusion that EDD supports CL outcomes; please revise the wording to make the distinction between descriptive results and conclusive claims clear.","section":"Section VI / Discussion"},{"comment":"The Default order is described as the 'OfficeDB default order,' but no citation is provided for this ordering; please cite the source or clarify how it was chosen.","section":"Section IV-D / Table IV"},{"comment":"Several references contain formatting errors, such as 'Y .-C. Hsu' and 'P. ¨Ogren'; please fix the spacing and accent glyphs throughout the bibliography.","section":"References"},{"comment":"The text in Section V-F states that semantic decomposition has a per-image latency of '~255 ms,' while Table IX reports a mean of 255.02 ms with SD 79.26; consider reporting the SD in the text or stating that it is the mean.","section":"Appendix / Table IX"},{"comment":"In Algorithm 1, the assignments h_E = phi_E(x_E; theta_E) and h_S = phi_S(x_S; theta_S) are written on a single line without an explicit separator; please format them as separate lines or add a semicolon for readability.","section":"Algorithm 1 / Step 8"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the lack of statistical support for the paper's central claim; the authors should either provide significance testing with per-fold analysis or substantially moderate the RQ1 conclusion. The evaluation datasets and some baselines come from the same research group, which is not circular reasoning but should be weighed when assessing novelty; the code release is a positive factor. If the statistical issues are resolved, the paper could be a useful contribution to continual learning for socially appropriate robot actions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely new: it frames socially appropriate robot action prediction as domain-incremental continual learning, and proposes a clean way to separate environmental from social cues via panoptic segmentation, feeding two branches and adding replay. The experimental scaffolding is solid — 5-fold CV, ablations, several baselines, public code, latency analysis. That is real value for the niche robotics/CL audience, and it is the first work I know of to do DIL for this task.\n\nThe soft spot is the central empirical claim. Table VII: EDD Bbox versus NoDec control is 0.78 vs 0.79 RMSE, 0.55 vs 0.53 PCC, 0.38 vs 0.36 CCC, all within one standard deviation of 5-fold noise. No significance test, no confidence intervals, no per-fold breakdown. Table VI is even more telling: single-branch CL (0.79/0.53/0.36) is essentially indistinguishable from dual-branch CL (0.78/0.55/0.38), while single-branch joint training (0.75/0.58) beats EDD CL. So the effect could easily be architecture or replay rather than explicit disentanglement. The paper's own RQ2 text says 'all methods perform relatively similarly,' yet the RQ1 conclusion asserts explicit separation supports CL outcomes. That is over-claiming. I agree with the stress-test: the central RQ1 claim is not yet empirically supported.\n\nMinor issues: the decomposition premise p(y|x)≈p(y|xE,xS) is untested under imperfect segmentation, and the authors admit this only as future work. Some baselines (DUCA, DARE++) were adapted with reduced resolution/batch size, which may put them at a disadvantage, though that is not the crux since the key control is NoDec. The datasets come from the same group, but that is fine for a benchmark.\n\nBottom line: this is a worthwhile empirical study with a plausible idea, but the headline claim needs statistical support or much more cautious language. I would send it to review — the formulation is useful and the artifacts are reproducible — but I would expect the authors to either provide significance tests/effect sizes or soften the conclusion.","headline":"Solid engineering and a genuinely new task formulation, but the paper's central claim about disentanglement rests on differences smaller than the noise; needs significance testing before that conclusion should be believed.","tokens_in":16796,"tokens_out":1979,"would_cite":false,"duration_ms":20697,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Explicitly separating environmental and social cues improves continual learning of socially appropriate robot actions.","keywords":["continual learning","domain-incremental learning","socially appropriate robot actions","environmental-social disentanglement","panoptic segmentation","dual-branch network","replay-based rehearsal","human-robot interaction"],"falsifier":"Take two scenes from different domains in which the panoptic segmentation model fails to detect one person, so that person remains visible in the environmental view. If EDD's predictions on those failure cases degrade no more than on correctly detected scenes, the social branch alone is carrying the signal; if they degrade sharply, the decomposition's benefit depends on near-perfect segmentation. A cleaner test: find or construct pairs of scenes with identical environmental masks and identical social masks but different human appropriateness ratings; any such pair would show that $p(y|x)\\approx p(y|x_E, x_S)$ is incomplete.","tokens_in":15754,"feed_emoji":"🤖","tokens_out":7954,"duration_ms":74455,"temperature":0.7,"pith_summary":"The paper claims that a robot learns socially appropriate actions across a sequence of new rooms more accurately, and forgets less, when each scene is first decomposed into an environment-only view and a social-agent-only view before learning begins. The proposed framework, EDD, builds these views with panoptic segmentation, processes them in two separate encoders, fuses the representations, and rehearses old examples from a small replay buffer. Across six indoor domains (a home and five office spaces) and nine robot actions, EDD reports the lowest prediction error and highest agreement with human appropriateness ratings among the tested continual-learning and vision-language baselines, and it stays close to the empirical joint-training upper bound. The result matters because robots deployed in real homes and offices cannot be pre-trained on every possible setting; they need to adapt as they move, without erasing what they already know.","feed_headline":"Separating people from rooms improves how robots learn room norms","feed_subtitle":"A dual-branch continual learner that treats environment and social cues separately stays near joint-training accuracy across six rooms.","key_machinery":"The load-bearing object is the semantic context decomposition, a pair of deterministic preprocessing transforms $x \\mapsto (x_E, x_S)$ that turn one scene into an environmental view and a social view using masks from a pretrained panoptic segmentation model. The environmental view blanks out detected agents and samples extra bounding-box occlusions from the current domain's mask pool; the social view keeps only agent silhouettes against a uniform background. Two separate encoders produce feature vectors $h_E$ and $h_S$, which are concatenated and passed to a regression head; a replay buffer mixes current-domain and past-domain samples during each training step. The structural assumption this machinery is built to satisfy is $p(y|x)\\approx p(y|x_E, x_S)$: the approximation that appropriateness depends on environment and social configuration separately.","core_discovery":"Domain-incremental continual learning of socially appropriate robot actions is improved by making the two sources of appropriateness information explicit at the input level. Instead of feeding full scene images to the network, EDD masks detected agents out of the environmental view, adding decoy bounding boxes to hide their locations, and masks the background out of the social view, then trains a dual-branch network whose fused representation predicts nine appropriateness scores. On the test domains, this decomposition reduces RMSE to 0.78 and raises Pearson and concordance correlation to 0.55 and 0.38 relative to single-branch and no-decomposition controls, while backward transfer stays near 0.02, meaning forgetting is small. The paper interprets these results as supporting the assumption that action appropriateness can be approximated as a function of environmental context and social configuration separately, with their combination recovered by concatenation.","pith_inferences":["Inference: A natural extension the paper does not run is a stress test with deliberately corrupted or missing segmentation masks; if EDD degrades gracefully as segmentation quality drops, the decomposition principle transfers to real robots with imperfect perception.","Inference: The equal per-domain quota in the replay buffer implicitly controls class balance; an order-aware replay policy that samples more from domains most similar to the current one could amplify the small curriculum effects reported.","Inference: The latency breakdown suggests that for real-time deployment, the bottleneck is the panoptic segmentation step at about 7.2 seconds per image, not the dual-branch network at about 20 milliseconds, so a faster segmentation backbone would make the approach practical.","Inference: Because the two branches are trained on visually distinct inputs, the fused representation offers a natural route to explainability: one could measure how much each branch changes the action scores, which the paper lists as future work but does not test."],"forward_implications":["A robot can keep adapting to new rooms with a replay buffer of about five percent of the training data, staying within a small margin of the joint-training upper bound in both error and correlation.","Using bounding-box masks for the environmental view outperforms silhouette masks, robot close-ups, and no decomposition, so how the separation is done matters more than the fact of separation.","Domain ordering has only modest effects on overall performance; the largest visible effect is that presenting the home domain last increases home-domain predictions, suggesting a mild curriculum effect.","Zero-shot vision-language models can reach competitive error rates but correlate weakly with human appropriateness ratings, indicating that foundation-model priors alone are not enough for this task."],"supporting_citations":[{"why":"Supplies the underlying dataset and the continual-learning formulation of social appropriateness that this paper extends to domain-incremental settings.","marker":"[10]"},{"why":"Provides the task-matched continual-learning baseline for socially appropriate actions and the practice of mean-aggregating human annotations.","marker":"[11]"},{"why":"Gives the prior non-continual approach and the evidence that foundation-model priors alone are insufficient for embodied social appropriateness.","marker":"[15]"},{"why":"Supplies the experience-replay mechanism that the rehearsal buffer is based on.","marker":"[25]"},{"why":"Provides the standard rehearsal training procedure with mixed current- and past-domain samples adopted for domain-incremental learning.","marker":"[38]"},{"why":"Gives a dedicated domain-incremental continual-learning method adapted here as a comparison baseline.","marker":"[39]"},{"why":"Gives another domain-incremental method used as a comparison baseline and achieves the strongest baseline correlations in the study.","marker":"[40]"},{"why":"Provides a zero-shot vision-language baseline whose competitive error but weak correlation supports the paper's critique of foundation-model priors.","marker":"[41]"},{"why":"Defines continual learning and catastrophic forgetting, motivating why the problem matters.","marker":"[3]"}],"fun_headline_variants":["Separating environment and social cues improves robot continual learning","Robot learns room norms better when rooms and people are decoupled","Dual-branch disentanglement reduces robot forgetting in new rooms","Explicit disentanglement boosts robot action learning across six rooms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands on the claim that hiding people to build the environmental view, and hiding the room to build the social view, throws away none of the information needed to judge whether an action is appropriate; if appropriateness depends on the joint interaction between who is present and where they are, this decomposition could destroy the very signal the model needs.","fun_headline_variants_meta":{"raw":{"variants":["Separating environment and social cues improves robot continual learning","Robot learns room norms better when rooms and people are decoupled","Dual-branch disentanglement reduces robot forgetting in new rooms","Explicit disentanglement boosts robot action learning across six rooms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000877,"raw_usage":{"total_tokens":3802,"prompt_tokens":966,"completion_tokens":2836,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":2750}},"tokens_in":582,"tokens_out":2836,"duration_ms":22656,"temperature":1.0,"reasoning_tokens":2750,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:44:22.998429+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two scenes from different domains in which the panoptic segmentation model fails to detect one person, so that person remains visible in the environmental view. If EDD's predictions on those failure cases degrade no more than on correctly detected scenes, the social branch alone is carrying the signal; if they degrade sharply, the decomposition's benefit depends on near-perfect segmentation. A cleaner test: find or construct pairs of scenes with identical environmental masks and identical social masks but different human appropriateness ratings; any such pair would show that $p(y|x)\\approx p(y|x_E, x_S)$ is incomplete.","supporting_citations":[{"cited_title":"Mind Your Manners! A Dataset and a Continual Learning Approach for Assessing Social Appropriateness of Robot Actions,","cited_arxiv_id":null,"evidence_quote":"Supplies the underlying dataset and the continual-learning formulation of social appropriateness that this paper extends to domain-incremental settings."},{"cited_title":"Grace: Generating socially appropriate robot actions leveraging llms and human expla- nations,","cited_arxiv_id":null,"evidence_quote":"Gives the prior non-continual approach and the evidence that foundation-model priors alone are insufficient for embodied social appropriateness."},{"cited_title":"Experience Replay for Continual Learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the experience-replay mechanism that the rehearsal buffer is based on."},{"cited_title":"Dual cognitive architecture: Incorporating biases and multi-memory systems for lifelong learning,","cited_arxiv_id":null,"evidence_quote":"Gives a dedicated domain-incremental continual-learning method adapted here as a comparison baseline."},{"cited_title":"Gradual divergence for seamless adaptation: a novel domain incremental learning method,","cited_arxiv_id":null,"evidence_quote":"Gives another domain-incremental method used as a comparison baseline and achieves the strongest baseline correlations in the study."},{"cited_title":"Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges,","cited_arxiv_id":null,"evidence_quote":"Defines continual learning and catastrophic forgetting, motivating why the problem matters."}],"review_version":1}