{"id":"467805d7-dd92-4fff-b906-fcd072b1cfd2","arxiv_id":"2607.17839","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Orientation-adaptive handovers reduced grasp delay for a wrench (2.55 vs 3.51 s, p=.003) and overall (2.72 vs 3.16 s, p=.014) against a static baseline.","lead":"A voice-driven robot handover system that rotates tools to match the receiver's hand orientation cut grasp delays for some asymmetric tools in a 15-person study. It uses an LLM to classify the tool request and MediaPipe to track the hand, then aligns the object to a per-tool 'handle-first' pose.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-tool grasp frames from an informal pilot (Sec. 2.3) are the main unvalidated link; with only the wrench significant, the 'asymmetric tools' claim may be a calibration artifact.","rationale":"The reader's weakest assumption is the empirical derivation of HTG; I agree this is the most load-bearing unvalidated element. The adaptive advantage is entirely mediated by HTG (Eq. 6), so if the pilot sample was small or unrepresentative, the significant overall and wrench results could reflect a favorable calibration rather than a robust method. The per-tool pattern (only wrench significant, screwdriver trending) is consistent with HTG being well-tuned for one tool but not others. The proposed test would directly check whether the chosen HTG is representative of the pilot users' natural grips, or whether it is an outlier. I do not see grounds to change the CONDITIONAL verdict; the study is promising but not conclusive until the calibration is validated.","tokens_in":11069,"tokens_out":12016,"duration_ms":129372,"concrete_test":"Provide the pilot-phase grasp-orientation data and, for each of the four tools, compute the circular 95% confidence interval of the natural grip orientations; test whether the selected HTG (Sec. 2.3, Fig. 3) lies inside that interval. If any chosen HTG is an outlier, the fixed per-tool frame is not representative of even the pilot users, and the adaptive benefit is not a generalizable property. If pilot data are unavailable, run a pre-registered replication with a separate calibration sample (N>=20) and a separate test sample (N>=20), re-deriving HTG from the calibration sample.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the adaptive system reduces grasp delay for asymmetric tools rests on the per-tool constant HTG used in Eq. (6). The paper states these values were 'empirically derived prior to the main experiment during an informal pilot phase,' but gives no details on the pilot group size, selection, or dispersion of their natural grips. Because the adaptive condition differs from the baseline only through these tool-specific HTG offsets, the observed grasp-delay advantage (overall 2.72 vs 3.16 s, p=.014; wrench 2.55 vs 3.51 s, p=.003) could reflect a favorable calibration for the wrench rather than a general receiver-centered principle. This concern is sharpened by the per-tool results in Table 3: only the wrench is significant, the screwdriver trend is not (p=.108), and the hammer and screw show no effect. That pattern is exactly what one would expect if HTG were well-tuned for the wrench but not for other asymmetric tools. Without an independent validation of HTG, the claim that the system generally reduces grasp delay for asymmetric tools is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a receiver-centered adaptive robot-to-human handover system implemented on a Franka Panda cobot with a RealSense camera. The system uses MediaPipe to estimate the receiver's 3D hand pose and an LLM (gpt-5-mini) restricted to classifying voice commands into one of four tools. The target object pose is computed as BTO = BTH · HTG (Eq. 6), where HTG is a constant per-tool grasp frame derived in an informal pilot; the approach motion is a validated cubic Bézier path with SLERP orientation interpolation. A counterbalanced within-subjects study (N=15) compared this adaptive condition with an object-agnostic baseline that tracks the hand position but keeps a constant orientation. The adaptive condition significantly reduced overall grasp delay (2.72 s vs 3.16 s; t(14)=−2.80, p=.014). By object, only the wrench showed a significant reduction (2.55 vs 3.51 s; p=.003); the screwdriver trended (p=.108), and hammer and screw showed no effect. Two of ten trust items favored the adaptive condition; NASA-TLX and blink rate did not differ significantly.","tokens_in":11302,"tokens_out":11311,"duration_ms":101418,"significance":"If the result holds, this is a useful integrated demonstration for HRI handover: voice-based intent recognition, real-time hand-orientation estimation, orientation-aware trajectory generation with IK/singularity validation, and an honest within-subjects evaluation with per-tool reporting. The manuscript provides exact test statistics, discloses the non-significant per-tool outcomes, and gives enough implementation detail (ROS2, Pinocchio, MediaPipe, FSM) to be reproduced; a demonstration video is provided. The bounded use of the LLM and deterministic safety validation are further strengths. The principal limitation is that the adaptive policy is parameterized by pilot-derived per-tool constants (§2.3); the experiment therefore evaluates a fitted policy rather than an independently derived prediction. With only one of four tools showing a significant benefit, the claimed generality of the grasp-delay effect for 'asymmetric tools' is not yet established, although a benefit for torque-requiring tools (wrench) is plausible and merits further testing.","major_comments":[{"comment":"§2.3, Eq. (6), Table 3: The adaptive condition's target pose is BTO = BTH · HTG, where HTG is a per-tool constant transformation 'empirically derived prior to the main experiment during an informal pilot phase.' No pilot size, selection, or grip-variability data are reported. Because the adaptive condition differs from the baseline only through these offsets, the grasp-delay advantage (overall p=.014; wrench p=.003) is evidence for this specific calibration, not yet for a general receiver-centered principle. The per-tool pattern (wrench p=.003; screwdriver p=.108; hammer p=.937; screw p=.628) is consistent with a well-calibrated wrench and poorer calibration elsewhere. Please report the pilot data, validate HTG out-of-sample, or restrict the claim to the calibration and tools tested.","section":"§2.3, Eq. (6); Table 3"},{"comment":"The abstract claims the adaptive system 'reduces the grasp delay for asymmetric tools,' but Table 3 shows a significant reduction only for the wrench; the screwdriver trend is not significant (p=.108) and hammer and screw show no effect (p=.937, p=.628). The overall effect (2.72 vs 3.16 s, p=.014) aggregates across four tools, two of which show no benefit. The conclusion's more hedged phrase 'for specific orientation-sensitive tools' is appropriate; the abstract should either be brought in line with that, or a principled, pre-specified definition of 'orientation-sensitive' tools should be provided and tested as a planned contrast.","section":"Abstract; §4 vs Table 3"},{"comment":"The text states that the non-significant NASA-TLX reduction (p=.221) 'constitutes a first indication of the superiority of the proposed handover policy' and that Fig. 5(b) shows workload and grasp delay 'co-varied coherently.' No correlation statistic is reported for Fig. 5(b), and a non-significant workload difference cannot be described as supporting superiority. Please report a formal correlation test (e.g., within-subject or per-condition) or reword the passage as a purely descriptive observation.","section":"§3.2, Fig. 5(b)"}],"minor_comments":[{"comment":"Ten trust items are tested with paired Wilcoxon tests and no multiple-comparison correction is applied (a note states this). The two significant items (p=.026, p=.034) would not survive Bonferroni correction; the per-tool analysis in Table 3 raises the same issue. Please label these as exploratory or apply/disclose a correction.","section":"Table 4"},{"comment":"The definition of grasp delay depends on detecting 'the user successfully grasping the object,' but the manuscript does not state how this event was detected (e.g., gripper force, camera event, manual annotation) nor the number of handover trials per participant/object/condition used in the participant means. Please specify these measurement details.","section":"§2.6"},{"comment":"The hybrid hand-pose method is described verbally; a small worked example or the exact landmark indices used for the MCP joints would improve reproducibility.","section":"§2.3"},{"comment":"No effect sizes are reported for the main comparisons; please add Cohen's d (or similar) for grasp delay, TLX, and blink rate to make the magnitudes interpretable.","section":"Global"},{"comment":"Typographical errors: 'guaranties' should be 'guarantees'; the preprint header contains broken spacing ('authenticate d version').","section":"§2.4"}],"recommendation":"major_revision","confidential_remarks":"To the editor: This is a workshop-proceedings submission (HFR 2026). For that venue, the paper is borderline-acceptable in its current form; the central issue is calibration. The within-subjects comparison is internally valid, but the pilot-derived HTG and the abstract's generalization overreach are load-bearing. I recommend major revision with the requirement that the authors either provide pilot details/validation or narrow the claims to the tested calibration and tools. The paper is not fatally flawed and the honest per-tool reporting is a point in its favor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honestly, this is a solid workshop-level paper with a clean, well-integrated system and a decent first user study. The genuinely new bit is the application of receiver-centered orientation adaptation to asymmetric mechanical tools, with a per-tool 'handle-first' grasp frame and a per-object grasp-delay measurement. That is not in the prior handover literature, which mostly sticks to generic objects or adaptive transport. The within-subjects design is counterbalanced, the metrics are standard, and Table 3 reports the per-object breakdown instead of hiding it. Credit for that transparency.\n\nThe main problem is that the abstract and conclusion claim the adaptive system reduces grasp delay for asymmetric tools, but the data only support that for the wrench (p = .003). The screwdriver is trending but not significant, the hammer and screw show nothing, and the overall effect is driven largely by the wrench. That is exactly the pattern you would expect if the per-tool grasp frame HTG, which was fitted during an informal pilot with an undescribed group, happened to be well-calibrated for the wrench and less so for the others. Since this HTG is the only difference between conditions, the circularity concern is real. The paper needs an independent validation of those frames, or a pre-registered replication with more participants, before the broad claim can stand.\n\nThe secondary findings are softer still. Two of ten trust items come out significant with no multiple-comparison correction, and the NASA-TLX result is explicitly non-significant but then described as a 'first indication' of superiority. That is overreading your own null result. These are minor in isolation, but they add to the impression that the framing is more enthusiastic than the evidence.\n\nNone of this kills the paper. The system works, the wrench result is credible, and the idea of per-tool grasp frames is worth pursuing. But the current write-up oversells the generality, and the methodology section needs to be much more open about the pilot details and the parameter-fitting risk.\n\nI would send this to peer review rather than desk-reject—it deserves a serious referee who can push on the calibration issue and the statistics. As it stands, it is a useful incremental contribution for people working on tool handover, but cite it as evidence of an approach, not as an established effect.","headline":"Useful incremental handover study, but the headline 'asymmetric tools' claim runs ahead of the data: only the wrench is significant, and the per-tool grasp frames were fitted in an undocumented pilot.","tokens_in":11828,"tokens_out":1345,"would_cite":true,"duration_ms":18364,"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":"An adaptive handover system that rotates tools to match the receiver's natural grip significantly reduces grasp delay for asymmetric tools, particularly the wrench, and improves trust in the robot's motion.","keywords":["Robot-to-Human Handover","Adaptive Handover","Human-Robot Interaction","Grasp Delay","Object Orientation","Hand Tracking","Trust","Collaborative Robots"],"falsifier":"Run the same within-subjects comparison with participants whose natural grip orientations differ from the pilot group, or measure the angle between the user's palm and the tool handle at the moment of grasp; if the adaptive advantage disappears or users re-adjust the tool in their hand, the claim that HTG encodes the ergonomic optimum would fail.","tokens_in":10909,"feed_emoji":"🤖","tokens_out":4019,"duration_ms":37884,"temperature":0.7,"pith_summary":"This paper tries to show that a robot handing a tool to a person should orient the tool to the person's hand, not just bring it to a fixed pose. It builds a system that tracks the receiver's hand, aligns the object's handle with that hand, and tests it against a static baseline in a 15-person study. The adaptive system shortens the time between the robot stopping and the person grasping the tool, especially for the wrench, and makes users feel the robot moves as expected. The result matters because tool handover is a frequent bottleneck in human-robot collaboration, and small differences in orientation can cause hesitancy or awkward wrist adjustments.","feed_headline":"Orienting tools to the user's grip cuts grasp delay 14%","feed_subtitle":"Handle-first presentation removes the wrist fumbling that slows tool exchange in shared workspaces.","key_machinery":"The central object is the constant grasp-frame transformation HTG, which encodes the ergonomically optimal orientation of each tool relative to the receiver's hand frame. It was derived from a small pilot group showing their natural grip for each tool. Multiplying the tracked hand pose by HTG yields the target object pose, so the robot presents the handle-first orientation aligned with the user's palm. The paper's mechanism is this one-step pose composition plus a cubic Bézier approach path that commits to a stable target once the hand is steady.","core_discovery":"The paper claims that receiver-centered orientation adaptation—rotating the end-effector so the object's grasp frame aligns with the user's hand orientation—reduces grasp delay for tools with a directional grip. In a within-subjects experiment with 15 volunteers, mean grasp delay fell from 3.16 s to 2.72 s overall, and from 3.51 s to 2.55 s for the wrench. The effect is attributed to eliminating the wrist adjustment users otherwise make when the tool arrives at a mismatched angle. Trust ratings also improved on two items: users were less worried because the robot moved as expected, and felt less discomfort from task complexity.","pith_inferences":["The paper leaves open whether the benefit generalizes beyond the four tested tools; a natural extension is to estimate the grasp frame online from the object's shape and the user's hand, rather than using a constant pre-set offset.","The null result for blink rate and workload, while not significant, hints that the adaptive advantage is specific to the physical grasp moment, not a general cognitive-load reduction.","If the HTG offsets are personalized to each user, the improvement may be even larger than the group average reported here.","The voice-activated LLM interface is incidental to the main claim; the core idea—orientation alignment—could be evaluated without it, which would isolate the effect."],"forward_implications":["If the adaptive orientation is adopted, tool handovers with asymmetric tools are likely to be faster and smoother in industrial settings.","Users may trust and accept robot partners more when the robot's motion matches their expectations.","The reduction in grasp delay suggests that ergonomic orientation is a larger factor than approach speed in handover fluency.","The approach could be applied to other objects with known grasp frames, reducing the need for the user to re-grip."],"fun_headline_variants":["Adaptive handover cuts tool grasp delay 14%","Robot reorients tools to your grip, cuts delay","Handle-first robot handover trims grasp delay","Receiver-centered handover cuts delay by 14%","14% faster tool handover with adaptive grip orientation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The per-tool grasp frames were set from a small informal pilot group rather than from the actual participants, so the adaptive condition might be tuned to a grip pattern that does not represent the general user.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive handover cuts tool grasp delay 14%","Robot reorients tools to your grip, cuts delay","Handle-first robot handover trims grasp delay","Receiver-centered handover cuts delay by 14%","14% faster tool handover with adaptive grip orientation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000789,"raw_usage":{"total_tokens":3273,"prompt_tokens":661,"completion_tokens":2612,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":405,"completion_tokens_details":{"reasoning_tokens":2549}},"tokens_in":405,"tokens_out":2612,"duration_ms":20593,"temperature":1.0,"reasoning_tokens":2549,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:49:07.518594+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same within-subjects comparison with participants whose natural grip orientations differ from the pilot group, or measure the angle between the user's palm and the tool handle at the moment of grasp; if the adaptive advantage disappears or users re-adjust the tool in their hand, the claim that HTG encodes the ergonomic optimum would fail.","supporting_citations":[],"review_version":1}