{"id":"9dc55eec-cc50-4fe4-abff-2629532add8b","arxiv_id":"2607.08857","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"AgenticFocus converts monocular human FPV videos into object-preserving mixed-reality humanoid demos with lower trajectory error and smoother wrist motion (SPARC −5.18) than Masquerade and Do as I Do.","lead":"AgenticFocus turns ordinary first-person human videos into robot training demos by restoring occluded objects, retargeting full hand motion to a humanoid, and compositing mixed-reality views with actions. It could cut the cost of collecting dexterous humanoid data without special capture gear.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Trajectory/SPARC gains do not establish robot-trainable supervision without policy or contact-geometry evidence.","rationale":"The reader correctly isolates the load-bearing gap: the paper’s strongest empirical results are kinematic (trajectory error + SPARC) while the claim is about policy-ready, object-preserving supervision for dexterous humanoids. Method §§2.2–2.4 describe object-template restoration and layered compositing as the mechanisms that make the data usable for contact-rich control, yet neither geometry fidelity nor any closed-loop policy result is measured. That is not a peripheral omission; it is the condition under which the abstract’s “robot-trainable demonstrations” claim would be substantiated. My concern is therefore the same as the reader’s weakest_assumption, only restated as the single hinge between the reported numbers and the title-level claim. A single policy-transfer experiment (or a quantitative contact/geometry ablation) would settle whether the concern lands. Until then CONDITIONAL remains the right verdict: accept the retargeting-fidelity contribution if claims are narrowed, or require policy evidence before accepting the full supervision claim. No stronger internal inconsistency or fabrication issue is present; the argument is simply incomplete relative to its stated goal.","tokens_in":8241,"tokens_out":655,"duration_ms":6306,"concrete_test":"Train the same visuomotor policy (e.g., a small ACT/Diffusion Policy) on AgenticFocus vs Masquerade/Do-as-I-Do outputs for a fixed set of EPIC-KITCHENS-style pick/place or wrap-around grasps, then evaluate success rate and contact error on a Unitree G1 + BrainCo hand in sim or real; if AgenticFocus does not improve success by a statistically clear margin, the “robot-trainable” claim does not hold on the current evidence.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that AgenticFocus turns ordinary monocular FPV into robot-trainable demonstrations for dexterous humanoid learning. The only quantitative support is lower mean 3D wrist/trajectory error and smoother SPARC (−5.18 vs −5.56/−6.05 on 75 episodes) versus Masquerade and Do as I Do (§3.2–3.3, Figs. 3–4). Those metrics evaluate retargeted wrist kinematics after camera-relative mapping (Eq. 2) and EMA smoothing (Eq. 3); they do not measure whether object-template reinsertion (§2.2) or layered full-hand + near-contact thumb compositing (§2.4) preserve contact geometry, depth ordering, or grasp-relevant visual features. No object-reconstruction error, contact-consistency score, or any downstream policy training/evaluation is reported; the Conclusion explicitly defers policy training. Thus the leap from “smoother retargeted wrist motion” to “robot-trainable supervision for dexterous humanoid learning” remains untested. Free parameters (β, T_offset, α) and monocular hand/object estimators further leave open whether the reported gains survive realistic embodiment and contact regimes.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"AgenticFocus proposes a Mixed Reality pipeline that converts ordinary monocular human FPV videos into synchronized humanoid training demonstrations. It isolates task-relevant objects (VLM + SAM2), restores occluded object geometry via an object-preserving inpainting mask Minpaint and template reinsertion, reconstructs full-hand motion (WiLoR/HaMeR-style), retargets it to Unitree G1 + BrainCo hands through camera-relative mapping (Eq. 2) and EMA-smoothed IK (Eq. 3), and composites via layered full-hand / near-contact thumb / object-template rendering. The paper reports lower mean 3D trajectory error and smoother wrist SPARC (−5.18 vs −5.56 Masquerade and −6.05 Do as I Do; 75 episodes, 95% CIs) under a shared protocol on EPIC-KITCHENS and internal clips, and claims the resulting focused visuals + actions/states constitute robot-trainable supervision for dexterous humanoid policies without specialized capture hardware.","tokens_in":8475,"tokens_out":1194,"duration_ms":24299,"significance":"If the pipeline truly yields usable visuomotor supervision at scale from ordinary FPV, it would lower a central data barrier for multi-fingered humanoids in household, assistive, and service settings, and would be a practical alternative to mocap gloves, stereo rigs, and scene-specific twins. Strengths include a constructive (non-end-to-end generative) decomposition, explicit object-preserving inpainting, full-hand rather than gripper-level retargeting, quantitative comparison against two external baselines with bootstrap CIs, and open acknowledgment that policy training is future work. The contribution is therefore of clear applied interest to the robotics and embodied-AI community, provided the leap from retargeting fidelity to trainability is either demonstrated or more carefully scoped.","major_comments":[{"comment":"The abstract, introduction, and contributions frame AgenticFocus as producing “robot-trainable demonstrations” and “supervision … suitable for training reactive humanoid visuomotor policies.” §§3.2–3.3 and Figs. 3–4 evaluate only intermediate retargeting fidelity (mean 3D trajectory error and wrist SPARC). No object-reconstruction error, contact-consistency, depth-ordering, or grasp-feature metric is reported for the template reinsertion (§2.2) or layered compositing (§2.4), and the Conclusion explicitly defers downstream policy training. Trajectory/SPARC gains therefore do not yet establish the central trainability claim; either add a minimal policy or contact-geometry evaluation, or substantially narrow the claim language to “retargeted visual-action pairs with improved kinematic fidelity.”","section":null},{"comment":"§2.3, Eqs. (2)–(3): retargeting depends on free parameters β (workspace scale), T_offset (virtual camera), α (EMA), and fixed wrist-orientation corrections, plus monocular hand/object estimators. The paper does not report sensitivity of trajectory error or SPARC to these choices, nor failure modes under heavy occlusion or embodiment mismatch. Without such analysis (or fixed public defaults and ablations), it is unclear whether the reported gains over Masquerade and Do as I Do are robust or pipeline-tuned.","section":null},{"comment":"§2.4 layered compositing (full articulated-hand pass + near-contact thumb pass + object template) is presented as preserving “plausible depth ordering” and “near-contact depth cues,” yet no quantitative or even systematic qualitative protocol (e.g., contact-frame occlusion consistency, multi-view check, or human preference study) is given. Because contact-rich geometry is load-bearing for the “dexterous” claim, this axiom needs either measurement or clearer limitation language.","section":null}],"minor_comments":[{"comment":"Affiliation line and author list contain spacing/encoding artifacts (“T echnology”, “T ara”, “†∗” style markers); clean for camera-ready.","section":null},{"comment":"Eq. (4) writes et =100 ∥…∥2 with an unusual “=100” placement; clarify units conversion and notation.","section":null},{"comment":"Fig. 3 and Fig. 4 captions are clear, but absolute trajectory-error magnitudes (cm) and the ground-truth source for “ground-truth trajectory” are not stated in the text of §3.2; add a short sentence on how GT is obtained for FPV clips.","section":null},{"comment":"Related work cites concurrent arXiv preprints (EgoEngine, Do as I Do, etc.) appropriately; ensure final versions and page numbers are updated if available at revision.","section":null},{"comment":"Index terms and abstract SPARC numbers match the body; keep consistent if any re-computation is done after ablations.","section":null}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid systems paper with a clear engineering contribution, but the title and abstract oversell “for Dexterous Humanoid Learning” relative to the experiments. I would accept after major revision if claims are scoped to retargeting quality or a lightweight policy transfer study is added. Fit for a robotics venue (ICRA/IROS/RSS-style or a journal special issue) is good; less so for a pure learning venue without policy results. No integrity concerns; novelty relative to Masquerade/EgoEngine is incremental but real (object-preserving template + full-hand camera-relative + layered contact compositing)."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing to know: this is a constructive pipeline paper that turns ordinary monocular FPV into mixed-reality humanoid demos with full-hand retargeting and object restoration under occlusion, and it beats two named baselines on wrist trajectory error and SPARC. The second thing: it never trains a policy, so the leap from “smoother retargeted wrists” to “robot-trainable supervision for dexterous learning” is still an assertion.\n\nWhat is actually new is the packaging, not a single algorithmic breakthrough. They protect the object during actor removal (Minpaint = hand ∩ ¬object), reinsert a clean object template, map full hand pose via camera-relative scaling/offset into Unitree G1 + BrainCo hands, and do layered compositing (full hand + near-contact thumb) so depth ordering near grasp is less broken than a flat overlay. That is a sensible answer to the occlusion + embodiment gap that Masquerade-style editing and gripper abstractions leave open. The empirical core is honest enough: 75 episodes, 95% CIs, SPARC −5.18 vs −5.56/−6.05, lower mean 3D error. Equations for camera-relative map and EMA smoothing are simple and readable. Citations land on the right neighbors (Masquerade, Do as I Do, EgoEngine, DexCap, SAM2, HaMeR/WiLoR).\n\nSoft spots, in proportion. The load-bearing gap is evaluation scope, not internal contradiction. Trajectory/SPARC only score wrist kinematics after free parameters (β, α, T_offset, wrist corrections). There is no object-geometry error, contact-consistency score, or any policy transfer. The conclusion admits this. Free parameters and monocular estimators are under-specified for replication. That weakens the abstract’s “robot-trainable” framing more than it weakens the retargeting result itself. Circularity is mild; they compare to external baselines on an external smoothness metric.\n\nWho it is for: people building humanoid data pipelines and cross-embodiment video editing who need a concrete monocular full-hand recipe. Not yet for someone who needs proof that the resulting demos train better policies. Math and data presentation are fine for a systems piece; no load-bearing formal claims to break.\n\nI would send it to peer review. Ask for either a narrowed claim (retargeting fidelity + object-preserving MR synthesis) or a small policy/contact study. Worth engaging if you care about FPV-to-humanoid data conversion; skip if you only care about closed-loop learning results.","headline":"Clean systems packaging of object-preserving full-hand FPV-to-humanoid synthesis; retargeting metrics look real, but the robot-trainable claim outruns the evidence.","tokens_in":9207,"tokens_out":636,"would_cite":true,"duration_ms":11130,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Ordinary first-person human videos can be turned into robot-trainable mixed-reality demos that keep object geometry and full-hand motion.","keywords":["Mixed Reality","Diminished Reality","Human-to-Robot Demonstration","Dexterous Manipulation","Humanoid Robots","Cross-Embodiment Retargeting","Egocentric Video","Object Restoration"],"falsifier":"Train the same humanoid visuomotor policy on AgenticFocus demos versus the two baselines on the same FPV clips and measure whether AgenticFocus yields reliably higher real-robot success on contact-rich grasps of small objects; or measure residual object-geometry and depth-ordering error at contact frames against ground-truth meshes.","tokens_in":9065,"feed_emoji":"🤖","tokens_out":662,"duration_ms":6375,"temperature":0.7,"pith_summary":"Human first-person video is abundant, but it is not robot data: hands hide the objects that matter most, cameras sit on human heads rather than robot torsos, and raw human motion is not paired with robot actions. AgenticFocus claims that a structured Mixed Reality pipeline can close those gaps without special capture hardware. It isolates the task object, restores a full object template after removing the human actor, reconstructs full-hand motion, retargets that motion into a camera-relative robot frame, and composites the robot hand and restored object in layers so depth near contact stays plausible. The output is a paired dataset of focused visuals plus synchronized robot actions and states. On trajectory reconstruction and wrist smoothness the method reports lower error and better SPARC scores than two cross-embodiment baselines, suggesting ordinary monocular FPV video can become usable supervision for multi-fingered humanoid policies.","feed_headline":"Human FPV video becomes robot demos without special hardware","feed_subtitle":"Object restoration and camera-relative full-hand retargeting cut trajectory error and smooth wrist motion","key_machinery":"AgenticFocus pipeline: object-preserving inpainting plus stable object-template reinsertion, camera-relative full-hand retargeting to a multi-fingered humanoid, and layered compositing (full-hand pass plus near-contact thumb pass) that keeps task objects and plausible contact depth while pairing visuals with robot actions and states.","core_discovery":"AgenticFocus shows that ordinary monocular human FPV videos can be converted into synchronized robot-trainable demonstrations by restoring occluded object geometry, retargeting full-hand motion through camera-relative alignment, and layered mixed-reality compositing, yielding lower mean 3D trajectory error and smoother wrist motion (SPARC −5.18 versus −5.56 and −6.05) than prior cross-embodiment baselines.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["FPV human video to humanoid demos via object restore and retarget","Ordinary monocular FPV becomes synchronized robot training data","Restored occluded objects from FPV cut robot trajectory error","Camera-relative hand retargeting yields smoother humanoid wrists","Mixed-reality compositing converts FPV video without special hardware"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That restoring a clean object template after actor removal and layering a full-hand render with a near-contact thumb pass is enough to keep contact geometry and depth ordering accurate for training multi-fingered robot policies.","fun_headline_variants_meta":{"raw":{"variants":["FPV human video to humanoid demos via object restore and retarget","Ordinary monocular FPV becomes synchronized robot training data","Restored occluded objects from FPV cut robot trajectory error","Camera-relative hand retargeting yields smoother humanoid wrists","Mixed-reality compositing converts FPV video without special hardware"]},"model":"grok-4.5","effort":"low","cost_usd":0.004642,"raw_usage":{"total_tokens":1263,"prompt_tokens":686,"num_sources_used":0,"completion_tokens":70,"cost_in_usd_ticks":46420000,"prompt_tokens_details":{"text_tokens":686,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":507,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":686,"tokens_out":70,"duration_ms":4340,"temperature":1.0,"reasoning_tokens":507,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T15:27:04.437763+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the same humanoid visuomotor policy on AgenticFocus demos versus the two baselines on the same FPV clips and measure whether AgenticFocus yields reliably higher real-robot success on contact-rich grasps of small objects; or measure residual object-geometry and depth-ordering error at contact frames against ground-truth meshes.","supporting_citations":[],"review_version":2}