{"id":"1495bda2-b5f5-4040-a00b-af7bbf146e2f","arxiv_id":"2611.13610","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A structured survey of Ego2Robot learning, organizing methods into four paradigms that extract different forms of supervision from egocentric human video for robot manipulation.","lead":"This paper surveys how robot manipulation policies can be learned from egocentric human videos (EHV) without recorded robot actions. It organizes the field into four learning paradigms and reviews the supporting datasets and open challenges.","discovery_kind":"review","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The taxonomy's four paradigms are not mutually exclusive, and the paper does not provide a systematic way to handle hybrid methods, which weakens the main organizational contribution.","rationale":"The reader's weakest assumption is that the taxonomy is both complete and mutually exclusive, and that the reviewed methods can be cleanly categorized into one paradigm. The paper acknowledges this overlap in Section 2.4.3 but does not provide a systematic way to handle hybrid methods. This is a legitimate concern, but it does not invalidate the paper's contribution. The paper is a survey, and its taxonomy is a useful organizational tool even if it is not perfectly exclusive. The paper acknowledges the limitations and provides a clear discussion of the overlaps. The verdict should remain ACCEPT because the paper substantively advances its field by providing a structured reference that will aid researchers in navigating and comparing approaches. The concern is noted but does not change the overall positive assessment.","tokens_in":46166,"tokens_out":600,"duration_ms":8477,"concrete_test":"Select 10 representative methods from the paper (e.g., R3M, LAPA, EgoMimic, Phantom, GR-2, DreamDojo, UniSkill, EgoVLA, MimicPlay, WARPED). For each, have two independent annotators assign it to one of the four paradigms, with the option to mark it as 'hybrid/unclear'. Compute the inter-annotator agreement (Cohen's kappa). If kappa < 0.6, the taxonomy is not reliably applicable, and the paper's main organizational contribution is weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the field can be organized into four distinct learning paradigms that differ in what information they extract from EHV and how they ground it in robot capabilities. However, the paper acknowledges in Section 2.4.3 that 'the two routes can be combined' and that 'many methods combine elements from multiple paradigms' (Section 2.4.3). The reader's weakest assumption correctly identifies this: the taxonomy is not robust if methods cannot be cleanly categorized. The paper does not provide a systematic way to handle hybrid methods, and the boundaries between paradigms are fuzzy. For example, latent action learning often uses explicit supervision (e.g., LAPA uses VQ-VAE with reconstruction loss, which is a form of explicit supervision), and representation pretraining is often combined with other paradigms (e.g., R3M uses temporal contrastive learning, which is a form of explicit supervision). If the taxonomy is not robust, the paper's main organizational contribution is weakened. The paper acknowledges this overlap in Section 2.4.3 but does not provide a systematic way to handle hybrid methods. This is a load-bearing concern because the taxonomy is the paper's primary contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This survey defines Ego2Robot learning as the process of learning or adapting robot manipulation policies from egocentric human video (EHV), and organizes the field into four learning paradigms: explicit embodiment transfer, explicit action and task supervision, egocentric video representation pretraining, and latent action and dynamics modeling. For each paradigm, the paper traces what information is extracted from EHV, how robot policies use it, and what additional data or grounding is required. It also reviews benchmark datasets spanning EHV, robot demonstrations, and synthetic or simulated resources, and closes with five open challenges: physical feasibility, agent coordination, in-context adaptation, controlled evaluation, and predictive control. The survey is explicitly scoped to methods that use EHV as a source of training data or task information, with robot demonstrations and simulated data treated as complementary grounding.","tokens_in":46511,"tokens_out":6238,"duration_ms":46798,"significance":"If the proposed taxonomy is accepted, this survey provides a genuinely useful organizing framework for a rapidly growing and fragmented literature. The paper's core criterion for the four paradigms—what information is derived from EHV, namely robot-format demonstrations, interpretable supervision, visual representations, or predictive models—is clear, and the paper consistently applies it throughout Sections 2.1-2.4. The survey is unusually careful about scope: it states that mixed-data methods are included only when their EHV use is documented, that inclusion does not imply attribution of gains to EHV, and that foundational or robot-only methods are discussed as context rather than evidence. The comparison with prior surveys in Table 1 and the dataset comparisons in Tables 2-4 are valuable resources. The paper also gives concrete, testable evaluation recommendations, particularly in Section 4.4, such as fixing architecture and training budgets when comparing EHV against non-egocentric video.","major_comments":[],"minor_comments":[{"comment":"The paper acknowledges that the two latent-action and world-action routes can be combined and that many methods blend paradigms, but it would help readers apply the taxonomy consistently if the authors stated explicitly that the four paradigms are not mutually exclusive and that a hybrid method is discussed under the paradigm corresponding to its primary EHV-derived output.","section":"Section 2.4.3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The sentence beginning 'Camera-frameposeaccuracyisinsufficientformotionretargeting' has missing spaces between words; please fix this and scan the manuscript for similar formatting errors.","section":"Section 2.1.1, Discussion","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The modality and annotation icon explanations appear garbled in the provided text, e.g., 'Modality icons denote♂layer-group depth, /adjus◎ monochrome video'; if this is an encoding artifact, it should be corrected in the published version.","section":"Table 2 caption","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The cross-reference to Section 4.2 for deployment-time use of interaction history is helpful, but 'interaction history' could be defined at first use to avoid ambiguity with the robot's proprioceptive or observation history.","section":"Section 2.2.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The 'Validation' heading with subcolumns 'Use' and 'Eval.' is somewhat confusing; renaming 'Use' to something like 'Role' or 'Training/Planning/Benchmark use' would clarify the intended distinction.","section":"Table 4","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"minor_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"The paper is a survey that organizes methods for learning robot manipulation from egocentric human video into four paradigms: explicit embodiment transfer, explicit action and task supervision, egocentric video representation pretraining, and latent action and dynamics modeling. The taxonomy is the main contribution, and it is genuinely useful for navigating a messy field. The paper does a good job of tracing what information each paradigm extracts from EHV, how it is grounded, and what additional data is required. The tables comparing datasets are comprehensive and well-structured. The authors are also unusually honest about limitations—they repeatedly note that paradigms overlap, that methods combine elements, and that the taxonomy is not mutually exclusive. They flag this in Section 2.4.3 and elsewhere. The stress-test concern about the taxonomy not being robust is real, but the paper already acknowledges it, and for a survey organizing a rapidly evolving field, a perfectly crisp taxonomy is probably impossible. The real soft spot is that the paper does not provide a systematic way to handle hybrid methods, which limits the taxonomy's utility as a prescriptive tool. But the authors are clear that this is a limitation, and the taxonomy still works well as a descriptive lens. The paper is well-written, the literature coverage is current (through mid-2026), and the open challenges section is thoughtful. It is not a methods paper, so novelty is in the organization and analysis, not in new results. That is fine for a survey. The paper deserves a serious referee and publication. I would cite it for the dataset comparison and the taxonomy framing. It is a solid reference for anyone working in this area.","headline":"A thorough, well-organized survey that delivers a useful taxonomy for a fast-moving area; the taxonomy is not airtight but the paper is honest about it, and the real value is in the structured comparison and the clear articulation of open challenges.","tokens_in":47008,"tokens_out":425,"duration_ms":7768},"pith_extraction":null,"created_at":"2026-11-11T13:48:58.557666+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":null,"supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}