{"id":"6dc9bc02-ee50-47f2-9c7b-dd59f25869f1","arxiv_id":"2603.08057","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A DINO-based eye-in-hand Switcher selects skill-part branches and flags anomalies in conditional PbD task graphs, reaching ~90% offline accuracy on user-gated decision windows from eight novice users.","lead":"See & Switch teaches robots conditional skills from human demos: a hand-camera Switcher picks the right next motion branch or asks for a new demo when the scene is novel. It matters because real tasks often need different motions depending on what is visible, and most PbD systems break without that.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Headline reliability claim rests on offline user-gated DS windows after filtering unobservable cases, not closed-loop automatic switching.","rationale":"The reader's weakest_assumption correctly identifies observability of the discriminative cue inside the DS window as the dominant failure mode the paper itself flags. That is the same soft spot: the offline user-gated protocol plus post-hoc removal of unobservable DS instances is what produces the ~90% numbers used to support \"reliable switching.\" I agree with the CONDITIONAL verdict and the medium correctness risk; there is no mathematical inconsistency, the engineering contribution (CIP-style graphs + DS-local DINO Switcher + multi-modal recovery) is real, and code/data are released. The single most load-bearing concern is therefore not a hidden error but the gap between the evaluation protocol that supports the strongest claim and the autonomous online setting the abstract advertises. Closing that gap with an automatic closed-loop re-evaluation would settle whether the claim holds without human DS timing and filtering.","tokens_in":14081,"tokens_out":555,"duration_ms":5119,"concrete_test":"Re-run the 576 rollouts (or a held-out subset) with fully automatic anomaly-triggered DS insertion and Switcher branch selection (no user gating or post-hoc DS filtering); report end-to-end Decision success and Task success rates. If automatic closed-loop Decision success falls substantially below the offline ~90% (e.g., below ~70–75%) or is dominated by timing/observability failures, the headline reliability claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that the vision Switcher demonstrates \"reliable switching based on visual input for conditional robot-skill programming\" is load-bearing on offline evaluation inside user-gated decision-state windows (Abstract; Sec. IV-B.2; Sec. V-B), not on automatic online closed-loop branch selection during full task execution. The paper explicitly uses user-gated DS timing \"to isolate the evaluation of the Switcher ... from temporal errors in automatic triggering\" and reports that filtering DS windows where the discriminative cue is not visible lifts branch accuracy from ~81.7% to 90.7% (Sec. V-B.1; red-marked instances in Figs. 9–11). Task-success numbers in Table II further use manual switching. Thus the strongest claim is true only under the joint conditions that (i) a human already places the DS window where the cue is visible and (ii) unobservable cases are removed post hoc. Without those conditions, the evidence does not yet establish reliable autonomous visual branching on the same novice data.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper presents See & Switch, a programming-by-demonstration framework that represents multi-variant manipulation tasks as graphs of skill parts linked by decision states (DS). A vision-based Switcher, built on frozen DINO features with lightweight heads, uses eye-in-hand images at DS context windows to (i) select among permitted successor skill parts and (ii) flag out-of-distribution scenes that require recovery demonstrations. Teaching is modality-agnostic (kinesthetic, joystick, hand gestures). The system is evaluated on three Robothon board tasks with 8 novice users; the Switcher is assessed offline on user-gated DS windows against SIFT/ORB and AEGP baselines, with a controlled class-growth study and a modality comparison of demonstration effort and replay success under manual switching. Code and data are released.","tokens_in":14313,"tokens_out":803,"duration_ms":12445,"significance":"If the claims hold under the stated evaluation protocol, the work is a solid incremental systems contribution to interactive PbD: it couples CIP-style conditional task graphs with a practical, DS-local vision classifier and a unified teaching API, and it provides a sizable novice-user real-robot dataset. Strengths include external baselines, a controlled scalability experiment (up to 8 classes), public code/data, and an honest discussion of observability failures. The result is useful for the community even if closed-loop automatic switching remains partially open, because offline DS-local visual branching on novice data is itself a non-trivial empirical target.","major_comments":[{"comment":"Abstract and Sec. V-B claim 'reliable switching based on visual input for conditional robot-skill programming,' but the primary evidence is offline classification inside user-gated DS windows (explicitly justified in Sec. IV-B.2 to isolate Switcher performance from automatic timing errors). Table II task-success rates further use manual switching. Please either (a) report closed-loop automatic branch-selection and full-task success under system-gated DS insertion on the same novice data, or (b) systematically soften abstract/conclusion language so the central claim matches the offline, user-gated protocol actually evaluated.","section":null},{"comment":"Sec. V-B.1 reports overall branch accuracy rising from 81.7% to 90.7% after removing DS windows with insufficient observability (red-marked instances in Figs. 9–11; ~40/78). The abstract headline (90.6%) appears to reflect the filtered figure. State clearly in abstract, results, and conclusion which number is primary, how many DS/windows/samples are excluded and by what objective rule, and report unfiltered accuracy as the main result with filtered as a diagnostic upper bound.","section":null},{"comment":"Numerical inconsistencies between abstract and body undermine reproducibility of the headline metrics: abstract cites ~900 rollouts, 90.6% branch accuracy, and >90% anomaly accuracy in 47 of 79 DS; body reports |R|=576, 90.7%, and ~43–44/78 DS (Figs. 9–11). Reconcile rollout counts (demonstrations × replays, filtered unsafe demos), DS counts, and accuracy figures throughout abstract, Sec. IV-B, and Sec. V.","section":null},{"comment":"Sec. III-D/E and Sec. V leave free parameters (context length e=10, percentile_keep=0.1, attn_keep, MIL training hyperparameters) largely unablated for sensitivity on the user-study set. Because anomaly thresholds and window placement directly affect both DS creation and reported accuracy, provide at least a short sensitivity analysis or fixed default protocol with justification so the 90% figures are not tied to unreported tuning.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean systems paper that does what it says: it takes CIP-style conditional task graphs and adds a vision-based Switcher so branch selection and OOD detection can run from eye-in-hand images, plus a modality-agnostic recovery layer (kinesthetic, joystick, gestures). That combination, plus a real novice user study and public code/data, is the actual contribution.\n\nWhat is new is not the graph model (Eiband et al.) or the anomaly idea (ILeSiA lineage, overlapping authors), but the DS-local frozen-DINO classifier that jointly does multiclass successor selection and novelty scoring, trained only on the permitted successors inside a short context window. They also ship a controlled class-growth experiment (2–8 states) where dinov2-small-attn stays above ~90% longer than SIFT or AEGP, and a user study with 8 novices, 3 tasks, 3 modalities, and hundreds of real-robot rollouts. The offline numbers are credible: ~81.7% branch accuracy overall, ~90.7% after dropping unobservable DS windows, and strong anomaly detection on a large fraction of DS instances. They are honest that limited observability is the dominant failure mode.\n\nThe soft spot is exactly the one the stress-test flags, and the paper itself is transparent about it. Headline “reliable switching” is measured offline inside user-gated DS windows chosen so the cue is visible; Table II task success uses manual switching. That isolates the visual classifier from timing and closed-loop errors, which is methodologically clean, but it means the strongest claim does not yet cover automatic online branching on the same data. Abstract/body count mismatches (576 vs ~900 rollouts, 47/79 vs ~43–44/78) are minor bookkeeping noise, not a soundness problem. Free parameters (window length, percentile_keep, attn_keep) are ordinary for this kind of system.\n\nMath is light and appropriate; citations are fair to the CIP/ILeSiA line and to DINO; no circularity. This is for people building interactive PbD and recovery tooling, not for theory. I would send it to peer review: the engineering is real, the data are real, and the limitation is stated. Engage if you care about visual branching in teachable task graphs; skip if you only want closed-loop autonomy claims.","headline":"Solid engineering extension of CIP with a DS-local DINO Switcher and real novice data; the reliability claim is real but scoped to offline user-gated windows, not closed-loop autonomy.","tokens_in":14996,"tokens_out":584,"would_cite":true,"duration_ms":6164,"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":"Eye-in-hand images can drive online branching in robot skills taught by non-experts, selecting the right recovery path or flagging new situations that need a demonstration.","keywords":["programming by demonstration","conditional skill graphs","decision states","vision-based branching","anomaly detection","eye-in-hand camera","DINO features","interactive robot teaching"],"falsifier":"Collect a new set of decision-state images in which the decisive object or state is deliberately kept outside the camera field of view, retrain the same Switcher, and check whether branch-selection accuracy collapses to chance while the rest of the pipeline remains unchanged.","tokens_in":14916,"feed_emoji":"🤖","tokens_out":659,"duration_ms":8133,"temperature":0.7,"pith_summary":"Programming robots by demonstration works for fixed motions but breaks when the world changes and a different sequence of actions is required. This paper argues that those conditional tasks can be taught as graphs of short skill parts linked at decision states, and that a vision-based Switcher can choose the right next part from the robot's own camera. At each decision state the Switcher both classifies among the known successor branches and detects when the scene is new, so a human can immediately teach a recovery behavior with kinesthetic guidance, a joystick, or hand gestures. On three board-level manipulation tasks taught by eight novices, the Switcher reaches roughly ninety percent branch-selection accuracy inside carefully timed decision windows and flags anomalies well on a large share of those windows. The result is a practical path from a single demonstration to a growing library of visual conditionals that non-experts can keep extending while the robot is running.","feed_headline":"Camera picks the right robot branch 90% of the time","feed_subtitle":"Novices teach conditional skills; a vision Switcher chooses recovery paths or asks for help","key_machinery":"The Switcher: a decision-state-local classifier built on frozen DINO patch embeddings that jointly scores among competing successor skill parts and raises an anomaly when the current image falls outside the known set, restricting decisions to a short temporal window and to the permitted branch set so that global scene understanding is unnecessary.","core_discovery":"Within user-gated decision-state windows, a DINO-feature Switcher that operates only on eye-in-hand images and only among the locally permitted successors can select the correct skill branch with up to about 90.7 percent accuracy and detect out-of-distribution scenes at greater than 90 percent accuracy on roughly half of the decision states collected from novice users, thereby making vision-driven conditional programming by demonstration reliable enough for real-robot use.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Vision Switcher hits 90% branch pick accuracy","Eye-in-hand images drive 90% robot skill branching","See & Switch: camera selects correct branches 90%","DINO Switcher achieves 90% skill branch accuracy","Robot vision picks right conditional skills 90%"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The visual cue that distinguishes one task variant from another (door open or closed, object present or missing) must actually appear in the eye-in-hand camera image inside the decision-state window; if it does not, correct branching is impossible from vision alone.","fun_headline_variants_meta":{"raw":{"variants":["Vision Switcher hits 90% branch pick accuracy","Eye-in-hand images drive 90% robot skill branching","See & Switch: camera selects correct branches 90%","DINO Switcher achieves 90% skill branch accuracy","Robot vision picks right conditional skills 90%"]},"model":"grok-4.5","effort":"low","cost_usd":0.00305,"raw_usage":{"total_tokens":1109,"prompt_tokens":805,"num_sources_used":0,"completion_tokens":63,"cost_in_usd_ticks":30500000,"prompt_tokens_details":{"text_tokens":805,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":241,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":805,"tokens_out":63,"duration_ms":2599,"temperature":1.0,"reasoning_tokens":241,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-15T12:51:34.508418+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Collect a new set of decision-state images in which the decisive object or state is deliberately kept outside the camera field of view, retrain the same Switcher, and check whether branch-selection accuracy collapses to chance while the rest of the pipeline remains unchanged.","supporting_citations":[],"review_version":1}