{"id":"fb4f1964-fbdf-4e76-ba18-1b456d00bdb9","arxiv_id":"2505.09771","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A 1-DoF three-finger gripper with V-shaped fingers and rigid fingernails grasps a wider range of objects more securely than flat two-finger grippers, with optional camera-based tactile force sensing.","lead":"The GET gripper adds a third fingertip to a standard one-degree-of-freedom parallel gripper, with two V-shaped fingers pressing against one opposing finger. It reports more secure grasps on 15 objects and 3 household tasks than flat two-finger grippers, and adds low-cost camera-based tactile force sensing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim of consistent outperformance is not yet statistically supported: object-level results use only 30 trials per finger and task results use 20, with no confidence intervals, significance tests, or per-operator variability.","rationale":"The reader's weakest_assumption is exactly the load-bearing point. The paper's headline comparison is an empirical claim about success rates, and the reported evidence consists of small, aggregated counts with no uncertainty quantification. This matters because the magnitudes of the differences in Figs. 9-10 are plausible but not necessarily statistically reliable at n=30 and n=20; a handful of trial outcomes could move a bar by 10-20 points. Inter-operator variation is unreported, and the 'secure' label depends on a subjective visual judgment during manual probing. None of this undermines the hardware design or the torque-resistance argument; it only means the central superiority claim should be conditional pending raw data or a larger, blinded evaluation. Since the reader already returned CONDITIONAL, no verdict change is needed.","tokens_in":11127,"tokens_out":9007,"duration_ms":94106,"concrete_test":"Release or reconstruct per-trial, per-operator data for all 15 objects and 3 tasks (the paper states code/CAD are on GitHub; trial logs would be a small addition). For each object/task, compare GET against the traditional flat-finger baseline with a two-sided Fisher exact test and compute Wilson 95% confidence intervals; additionally fit a logistic regression with operator as a random effect. If fewer than 15 of 15 objects and 3 of 3 tasks show a significant advantage (or if any object-level CI overlaps the baseline estimate), revise the abstract's 'consistently outperformed' to a qualified statement such as 'tended to outperform' or 'outperformed on most objects'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-B reports success for each of 15 objects as 10 trials × 3 operators = 30 Bernoulli trials per finger, aggregated without confidence intervals or significance tests; Section IV-C does the same for 20 attempts per manipulation task. At n=30, a 20-percentage-point gap (80% vs 60%) has a two-proportion z of about 1.7 and p≈0.08, so many of the visually large bars in Figs. 9-10 are within sampling noise. The 'secure' criterion for large objects is also operator-judged: no visual slip while probing 'up to 3 N' in any direction, with no blinding or inter-rater reliability measure. If a few trials flip, several object-level advantages disappear. The design rationale and torque analysis are plausible, but the empirical assertion 'consistently outperformed standard flat fingers' is not quantified enough to sustain ACCEPT-level confidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces GET, a 1-DoF, three-fingered gripper attachment for standard parallel-jaw actuators. The design pairs two V-shaped fingers against one opposing finger, adds rigid fingernails for small-object acquisition, and is parametrically scaled for ALOHA, Franka Panda, and UMI embodiments. A camera-based tactile sensor in the single finger estimates normal force from difference images via a ResNet-18 network. The authors present a geometric/torque argument for why three contacts improve grasp security, then report experiments: force-estimation validation (Table I), grasp-success comparisons on 15 objects (Fig. 9), and three teleoperated manipulation tasks (Fig. 10), with GET compared against ALOHA ViperX fingers and a custom 'traditional' flat-finger design. The central claim is that GET 'consistently outperformed standard flat fingers' in these experiments.","tokens_in":11318,"tokens_out":5647,"duration_ms":56504,"significance":"If the empirical claim is substantiated, this is a useful, low-cost, retrofittable gripper design that improves robustness without adding actuated degrees of freedom, which matters for learning-from-demonstration pipelines and in-the-wild data collection. The open-source release of CAD files and the integration of proximity-plus-tactile imaging are strong practical contributions. The force-estimation network is trained on held-out validation and unseen-object data rather than fitted to the reported experiments, which is a strength. The main unresolved issue is the statistical support for the central comparative claim, which currently rests on aggregated Bernoulli trials without uncertainty quantification.","major_comments":[{"comment":"The object-level comparison rests on 30 trials per finger per object (10 trials x 3 operators) aggregated into bar heights with no confidence intervals, significance tests, or per-operator breakdown. At n = 30, a 20-percentage-point difference (e.g., 80% vs 60%) is not statistically significant (two-proportion z about 1.7, p about 0.08), so many of the visually large differences in Fig. 9 are within sampling noise. Additionally, the 'secure' criterion for large objects is operator-judged (no visual slip under manually applied probes 'up to 3 N') with no blinding or inter-rater reliability measure. Please report raw success counts per operator, exact confidence intervals (e.g., Clopper-Pearson), and appropriate significance tests (e.g., Fisher exact or McNemar for paired trials), and either blind the slip assessment or demonstrate inter-rater agreement. Without these, the central claim of consistent superiority is not supported.","section":"Section IV-B, Fig. 9"},{"comment":"The task-level claim that 'GET fingers more successfully completed all tasks' is based on 20 attempts per gripper per task, again with no confidence intervals, significance testing, or per-operator variability. The completion-time metric is the average over successful trials, but the number of successful trials is very small for some baselines (the figure labels suggest as few as 1-2 successful trials for some conditions), making the reported means unstable. Please provide the number of successes per condition, confidence intervals for success rates, per-operator results, and a clearly defined pass/fail criterion for each of the three tasks.","section":"Section IV-C, Fig. 10"},{"comment":"The torque-capacity formulas |tau_x| <= Fmax(w + L/2) and |tau_z| <= mu Fmax(w + L/2) assume that each of the three contacts can independently exert the full actuation force Fmax and that L is a fixed geometric parameter. In a 1-DoF parallel jaw, the actuation force is shared among the contacts, and the normal-force distribution depends on object geometry and compliance; the formulas should be derived from an actuator-level force balance or explicitly qualified as an upper bound. Also, because the V-shape makes L vary along the finger, the text should specify where L is evaluated for a given grasp. The qualitative point that a third contact can increase torque resistance is plausible, but the specific expressions are not justified by the text as written.","section":"Section III-A1, Fig. 3b"},{"comment":"The abstract states that the network was trained 'with an average validation error of 1.3 N', but Table I reports a combined validation RMSE of 1.34 N and unseen-object errors of 1.40, 2.65, and 2.83 N (average 2.29 N). Please state the metric explicitly (RMSE versus MAE) and clearly distinguish validation-set performance from unseen-object generalization; the current wording overstates generalization accuracy. In addition, clarify whether the 20% validation split was object-disjoint or a random frame split, since random splits of repeated indentations can optimistically bias the validation RMSE.","section":"Section IV-A, Table I and abstract"}],"minor_comments":[{"comment":"The term 'self-similarity' is used as an analogy, but no actual self-similar structure or scaling law is defined; consider renaming this 'parametric scaling' or providing a concrete scaling argument.","section":"Section III-A2"},{"comment":"There are several typos, including 'desgins' in Fig. 6 and 'compatability' in Section III-A2; please correct them.","section":"General"},{"comment":"Fig. 10 reports average completion time over successful trials without any measure of spread or the number of successes; please add error bars or a table with per-condition success counts and completion times.","section":"Section IV-C"},{"comment":"The paper states that designs are 'available on GitHub' but provides no repository URL; please include the link in the final version.","section":"Section IV-A"},{"comment":"The claim of 'low cost (less than $100)' would be more credible with a brief bill of materials or component cost breakdown.","section":"Section I, contributions"},{"comment":"For the unseen-object rows, please state explicitly that these 800 images per object were not used in training; as written, the reader cannot determine whether these entries are part of the training set.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The core weakness is the statistical support for the empirical claim. The authors should be asked to provide raw trial data and reanalyze with confidence intervals and significance tests, or substantially soften the 'consistently outperformed' wording. The design rationale and open-hardware contribution are solid, and the paper appears within scope for a robotics venue; this is fixable within a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The GET gripper is a genuinely useful design idea: three narrow fingers on a standard parallel-jaw actuator, arranged two-against-one with a V-shaped pair, plus rigid fingernails for small objects and an optional camera-based tactile sensor. The combination is new as far as the cited literature goes, and the paper does the right things around it—parametric resizing for different arms, open CAD and code, a clear fabrication guide, and a force-estimation network trained on held-out objects. The torque analysis in Fig. 3b is simplified but honestly stated, and it works as motivation for why a third finger helps resist disturbances about the tool axis. Credit where due: this is a practical, low-cost upgrade for the large installed base of ALOHA-style 1-DoF grippers, and the authors clearly know the hardware space.\n\nThe soft spot is exactly what the stress-test note says: the empirical case for 'consistently outperformed' is not yet quantified. Fifteen objects times 10 trials per operator times 3 operators gives 30 Bernoulli trials per finger per object. At that n, a 20-percentage-point gap is within sampling noise (z≈1.7, p≈0.08). The manipulation tasks use only 20 attempts per condition. No confidence intervals, no significance tests, no per-operator breakdown. The 'secure' criterion for large objects is also operator-judged—no visual slip under hand-probed forces up to 3 N—with no blinding or inter-rater check. If a few trials flip, several apparent advantages disappear. This is the central load-bearing weakness, and it is fixable: more trials, a pre-registered protocol, and at least a per-operator table.\n\nTwo smaller notes. The self-similarity/scaling claim is asserted and illustrated with a ×1.5 Panda version, but never tested quantitatively; that is fine as future work, but the paper should not lean on it. And the force-sensing validation numbers (1.3 N average, up to 2.83 N on unseen objects) are honestly reported, though the 2.65–2.83 N errors on unseen objects suggest the model is not yet a precision force sensor.\n\nOverall: the design contribution is solid, the engineering is reproducible, and the evaluation weakness is addressable rather than fatal. Citation pattern looks fair, including prior art on fingernails and tactile sensing. I'd send this to peer review—a good reviewer will push for proper statistics, and the paper will be better for it. I'd probably also cite it if I worked on low-cost manipulation hardware.","headline":"A clever, well-documented 1-DoF gripper design with a plausible torque advantage, but the empirical 'consistent outperformance' claim needs more statistical support before I'd trust it.","tokens_in":745,"tokens_out":847,"would_cite":true,"duration_ms":20595,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A one-degree-of-freedom gripper with three tapered fingers can hold objects more securely than the standard two flat fingers used across robotics.","keywords":["parallel-jaw gripper","three-finger grasping","robotic manipulation","tactile sensing","force estimation","teleoperation","grasp stability","self-similar design"],"falsifier":"Recompute the grasp results per operator and per object with binomial confidence intervals; if the flat-finger and GET success intervals overlap for a majority of the 15 objects, the claimed consistent advantage is not supported. A sharper test would instrument a fixed grasp with a six-axis force-torque sensor and measure the maximum disturbance torque about the grasping axis before slip for GET and flat fingers on identical cylindrical handles; the geometry predicts GET's limit should grow with finger separation $L$, so a failure to find that scaling would refute the mechanism.","tokens_in":10931,"feed_emoji":"🤏","tokens_out":6553,"duration_ms":64728,"temperature":0.7,"pith_summary":"This paper aims to show that a simple hardware change—three narrow, tapered fingers on a standard one-degree-of-freedom parallel jaw—makes robotic grasps more secure than the usual two flat fingers. The design is a two-against-one layout: two fingers converge into a V, and a third opposes them, so grasped objects meet three contact patches instead of two. From a 2D statics argument, the third contact gives a lever arm that resists twisting forces about the grasping axis and along the finger direction, something flat jaws cannot do without very wide fingers. The authors support the claim with teleoperated trials on 15 objects and three tool-use tasks, plus a camera-based tactile finger that estimates normal force to about 1.3 N validation error. If the results hold, any parallel-jaw robot could gain more reliable grasping at low cost and without changing its actuator.","feed_headline":"Three tapered fingers beat flat jaws at secure grasping","feed_subtitle":"A 1-DoF gripper add-on adds torque resistance, small-object nails, and camera-based touch for under $100.","key_machinery":"The load-bearing mechanism is the two-against-one, V-shaped finger geometry. The two V fingers are pitched inward by a small angle so they can flatten into interdigitation and still grasp very thin objects, while the varying separation $L$ along their length lets objects of different sizes contact at different positions—small objects near the tip, large objects near the base where the lever arm is longest. Soft silicone gel pads deform to the object's local shape, enlarging contact patches and adding elastic resistance to slip; rigid fingernails at the tips handle prying small objects. For tactile sensing, the single finger's backing is left optically clear, so an external high-dynamic-range camera sees both the gel deformation and nearby objects in the same image, and a neural network maps difference images to normal force.","core_discovery":"On its own terms, the paper's central discovery is that replacing the two flat fingers of a parallel jaw gripper with a three-finger, V-shaped pair-plus-opposing configuration turns a 1-DoF gripper into a device that can form three-point, geometry-conforming grasps. The authors derive that the maximum disturbance torque a two-finger grasp can resist scales with finger width $w$, while the three-finger grasp resists torque approximately in proportion to the lever arm $L$ between fingers, giving an advantage that grows with object size and with distance between contacts. They also find that rigid fingernails at the fingertips let the gripper pry small flat objects off a table, and that a camera mounted behind a soft gel pad, with colored LED lighting, yields tactile images from which a convolutional network estimates normal contact force with roughly 1.3 N validation error across varied geometries. In experiments, the GET fingers grasped small objects and securely held large ones more often than the flat-finger baselines, and completed three teleoperated tasks—hammering, spreading, and small-part assembly—with higher success and shorter completion times.","pith_inferences":["A natural extension the paper does not pursue is an autonomous grasp planner that uses the varying lever arm $L$ explicitly—for example, adjusting grasp position along the finger to optimize torque resistance for a given object size; the geometry suggests this is a tunable quantity rather than a fixed property.","The statics comparison predicts the three-finger advantage grows with the lever arm $L$, so one could expect diminishing returns once $L$ exceeds the object width; this is testable by measuring slip torque on scaled finger versions.","Because the experiments used human teleoperators, the consistent-outperformance claim is about what a skilled operator can achieve; whether learned policies trained on GET data retain the same advantage is an open consequence, not demonstrated here."],"forward_implications":["If the central claim is correct, any existing parallel-jaw robot can adopt the finger geometry without changing its actuator, motor, or control interface, so the grasp improvements transfer to systems already deployed for teleoperation and data collection.","The parametric sizing means the same design can be scaled for arms of different payloads and workspaces; the paper demonstrates versions for three different robot platforms.","The tactile finger adds a force-sensing channel at roughly 1.3 N validation error without requiring a wrist force sensor, so learning-based policies could use force feedback from the same image stream that already records contact and proximity.","Rigid fingernails extend the range of graspable objects to small, flat, or clustered items like coins, paper clips, and rubber parts, which flat fingers typically cannot pick from a table or from clutter."],"supporting_citations":[{"why":"Supplies the teleoperation hardware and data-collection platform used in all grasp and task experiments.","marker":"[1]"},{"why":"Provides the enhanced gripper variant onto which the three finger designs were mounted for comparison.","marker":"[6]"},{"why":"Background theory for why two antipodal contact patches limit torque resistance, motivating the three-contact design.","marker":"[7]"},{"why":"Inspires the compliant, thickness-varying finger structure that increases contact area and conformity.","marker":"[11]"},{"why":"Earlier work combining compliant fin-ray-style fingers with tactile sensing, informing the gel pad design.","marker":"[12]"},{"why":"Introduces camera-based tactile sensing through a soft gel pad, the basis for the sensor in the single finger.","marker":"[22]"},{"why":"Supplies the neural network architecture adapted for normal-force estimation from tactile difference images.","marker":"[34]"},{"why":"Self-similarity principle used to justify the size-scaling and interchangeability of the finger design.","marker":"[30]"}],"fun_headline_variants":["1-DoF triple-finger gripper beats flat jaws at secure grasping","V-shaped fingers: 1-DoF gripper conforms to objects","Three-finger add-on improves 1-DoF gripper's grasp stability","Tactile 3-finger gripper estimates force from camera images","In 15-object tests, V-finger gripper outperformed flat jaws"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that 30 trials per finger design across three teleoperators, with 'secure' judged by visually checking for slip under hand-applied probes of up to 3 N and no confidence intervals or significance tests, is enough evidence that the three-finger design reliably beats flat fingers.","fun_headline_variants_meta":{"raw":{"variants":["1-DoF triple-finger gripper beats flat jaws at secure grasping","V-shaped fingers: 1-DoF gripper conforms to objects","Three-finger add-on improves 1-DoF gripper's grasp stability","Tactile 3-finger gripper estimates force from camera images","In 15-object tests, V-finger gripper outperformed flat jaws"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000832,"raw_usage":{"total_tokens":3672,"prompt_tokens":1023,"completion_tokens":2649,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":2562}},"tokens_in":639,"tokens_out":2649,"duration_ms":20254,"temperature":1.0,"reasoning_tokens":2562,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:24:54.601299+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the grasp results per operator and per object with binomial confidence intervals; if the flat-finger and GET success intervals overlap for a majority of the 15 objects, the claimed consistent advantage is not supported. A sharper test would instrument a fixed grasp with a six-axis force-torque sensor and measure the maximum disturbance torque about the grasping axis before slip for GET and flat fingers on identical cylindrical handles; the geometry predicts GET's limit should grow with finger separation $L$, so a failure to find that scaling would refute the mechanism.","supporting_citations":[{"cited_title":"Grasping,","cited_arxiv_id":null,"evidence_quote":"Background theory for why two antipodal contact patches limit torque resistance, motivating the three-contact design."},{"cited_title":"Fin ray® effect inspired soft robotic gripper: From the robosoft grand challenge toward optimization,","cited_arxiv_id":null,"evidence_quote":"Inspires the compliant, thickness-varying finger structure that increases contact area and conformity."},{"cited_title":"Gelsight fin ray: Incorporating tactile sensing into a soft compliant robotic gripper,","cited_arxiv_id":null,"evidence_quote":"Earlier work combining compliant fin-ray-style fingers with tactile sensing, informing the gel pad design."},{"cited_title":"Allsight: A low-cost and high-resolution round tactile sensor with zero-shot learning capability,","cited_arxiv_id":null,"evidence_quote":"Supplies the neural network architecture adapted for normal-force estimation from tactile difference images."},{"cited_title":"Fractals and self similarity,","cited_arxiv_id":null,"evidence_quote":"Self-similarity principle used to justify the size-scaling and interchangeability of the finger design."}],"review_version":1}