{"id":"1dfed3e4-57be-4025-a725-b4e3eca8c4f8","arxiv_id":"2509.04324","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"OVGrasp integrates YOLO-World open-vocabulary detection, depth-based target selection, and speech-triggered release to control a cable-driven soft exoskeleton for assistive grasping.","lead":"This paper presents OVGrasp, a soft hand exoskeleton controller that uses open-vocabulary object detection and voice commands to trigger grasps and releases. The authors report a grasping ability score of about 87% across ten healthy participants and claim better hand kinematics with an egocentric camera.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Finger-ROM evidence rests on an invalid pose estimator and an uncontrolled camera-viewpoint confound, undermining the kinematic-alignment claim.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the finger-joint ROM comparison is not trustworthy. The additional Table 3 inversion and Table 1 pinch-grip deficit are real but secondary; the kinematic claim is essential to the paper's contribution and is explicitly part of the strongest claim. The described pose-estimation procedure is not internally plausible because 11k Hands lacks joint annotations, and even a valid pose model would need viewpoint-invariance or per-view validation to compare head-mounted and hand-mounted cameras. Without that, the ROM bars and significance stars in Figure 4 are not evidence for the egocentric advantage. A motion-capture validation would settle the issue directly. Therefore the reader's REJECT verdict stands; no change.","tokens_in":14846,"tokens_out":6574,"duration_ms":57842,"concrete_test":"Repeat the protocol in §4.3 with ten participants using a marker-based motion-capture system (e.g., Vicon/OptiTrack) or instrumented goniometers to obtain ground-truth DIP/PIP/MCP angles in all three conditions (exo off, eye-in-hand exo on, egocentric exo on). Test whether the egocentric-vs-eye-in-hand difference in ROM relative to exo-off remains significant and of the same magnitude; if the effect disappears, the kinematic-alignment claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 states that finger ROM (DIP/PIP/MCP) was measured by fine-tuning YOLO11 on the 11k Hands Dataset over 100 epochs. The 11k Hands dataset contains hand images for detection/segmentation and does not provide MCP/PIP/DIP joint keypoint annotations, so this pipeline cannot yield joint angles as described. No validation of the estimated angles against motion capture or goniometry is provided. Moreover, the two comparison conditions (eye-in-hand vs egocentric) mount the camera in physically different locations (under the hand vs on the head), so viewpoint-dependent pose-estimation bias is confounded with any true kinematic difference. The ROM differences in Figure 4 thus cannot be attributed to the control framework; they may be artifacts of the pose estimator or camera mounting. This directly undermines the abstract's central claim of 'improved kinematic alignment with natural hand motion.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"OVGrasp is a hierarchical control framework for a cable-driven soft hand exoskeleton that combines YOLO-World open-vocabulary detection with depth- and speech-based intent fusion and PID low-level control. The paper reports hardware integration, a 15-object, 10-participant evaluation using the Grasping Ability Score (GAS), zero-shot detection results, and finger joint ROM comparisons between eye-in-hand and egocentric camera configurations. The central claims are that OVGrasp outperforms push-button and force-sensing baselines and improves kinematic alignment with natural hand motion.","tokens_in":15100,"tokens_out":5008,"duration_ms":42110,"significance":"If the results held, the paper would make a useful contribution to wearable assistive robotics by demonstrating an open-vocabulary multimodal controller and a quantitative comparison of egocentric vs eye-in-hand sensing for grasp assistance. It also presents a complete system architecture and uses a standardized GAS protocol. However, the experimental evidence contains major internal inconsistencies and a kinematics measurement pipeline that is not supported by the cited dataset, so the headline claims are not currently established.","major_comments":[{"comment":"The abstract claims OVGrasp 'outperforms state-of-the-art baselines,' but Table 1 shows that for pinch grasps the proposed method (GAS 87.65±0.42) is below both push-button (92.00±0.95) and force-sensing (93.00±1.17). Since pinch is one of the three evaluated grasp types, the blanket claim is unsupported. The authors should either qualify the claim to the overall average or to specific grasp types, and report pairwise significance tests; as presented, the standard deviations overlap in a way that makes the claimed advantage unclear.","section":"§5.1, Table 1"},{"comment":"The table and text directly contradict each other. Table 3 reports YOLO with OVD has APunseen=0.03, mAP=0.23 while YOLO without OVD has APunseen=0.86, mAP=0.93; the text states the exact opposite ('rises sharply from 0.03 to 0.86'). Additionally, the text mentions '93.3% compared to 33.3%' which appears in no table entry. As printed, the ablation either refutes the zero-shot generalization claim or is mislabeled; in either case the conclusion cannot be assessed. The authors must correct the table/text and reconcile the numbers before the open-vocabulary contribution can be evaluated.","section":"§5.3, Table 3 and surrounding text"},{"comment":"The finger-ROM analysis is not valid. The 11k Hands dataset is a hand segmentation and detection dataset; it does not contain MCP/PIP/DIP joint keypoint annotations, so fine-tuning YOLO11 on it cannot produce joint angles as described. No validation against motion capture or goniometry is provided. Moreover, the two exoskeleton 'on' conditions place the camera in different physical locations (under-hand vs head-mounted), so viewpoint-dependent pose-estimation bias is confounded with any real kinematic effect. Figure 4 therefore cannot support the claimed 'improved kinematic alignment with natural hand motion.'","section":"§4.3, §5.2, Fig. 4"}],"minor_comments":[{"comment":"The phrase 'Weight YOLO-World-X' appears to have a typo; it should state 'with weights of YOLO-World-X' or similar.","section":"§3.2.1"},{"comment":"The grasp-point graph G=(V,E) introduces edges E that are never defined or used in the algorithm; either remove the edge set or describe how edges contribute to decision-making.","section":"§3.2.2"},{"comment":"The abstract's overall GAS 87.00±2.46 does not correspond to any single row in Table 1; specify how this aggregate is computed across the three grasp types and ten participants.","section":"§5.1"},{"comment":"The distinction between 'seen' and 'unseen' objects is tied to YOLO-World's pretraining data. Clarify that 'seen' means categories present in that pretraining set, and confirm that the 'unseen' objects are genuinely absent from the pretraining vocabulary.","section":"§4.1"},{"comment":"Use consistent units for AP/mAP (fractions vs percentages). Mixing values like 0.86 and 93.3 makes the table hard to read and can hide inconsistencies.","section":"Table 3"},{"comment":"The DepthRegister function is not defined; state the spatial registration method used to align depth and RGB frames.","section":"Algorithm 1"}],"recommendation":"reject","confidential_remarks":"The internal contradiction in Table 3 is severe enough that I recommend the editor request the raw detection outputs and the original ablation logs during any revision. The row/column labels appear as if they may have been swapped, but even after a swap, the seen-class AP dropping from 1.0 to 0.43 with OVD is unusual and needs explanation. The kinematic analysis would require a completely new experiment with a validated hand keypoint estimator or motion capture; it is not a small local fix."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the new thing here is real—YOLO-World plus depth and speech as a grasp/release controller on a soft exoskeleton, with a systematic egocentric vs eye-in-hand comparison on the same hardware. That comparison hasn't been done before, and the GAS protocol with 10 participants, 15 objects, and three grasp types is concrete and reproducible. The hardware documentation is careful, and the release-intent via voice plus the queue-based grasp trigger are sensible engineering choices.\n\nThe problems are in the evidence. First, Table 3 has the rows reversed relative to the text. The text says the OVD-enabled detector reaches AP 93.3% and APunseen 0.86, but those numbers sit in the \"YOLO without OVD\" row; the \"with OVD\" row shows APunseen 0.03. Either the table labels or the text is wrong, and the section's argument depends on the text's reading. That's not a minor typo—the ablation as printed contradicts its own conclusion. Second, the abstract claims OVGrasp \"outperforming state-of-the-art baselines,\" but Table 1 shows it below both push-button and force-sensing on pinch GAS (87.65 vs 92.00 and 93.00). The body is more careful, crediting only spherical and cylindrical, but the abstract overclaims. Third, and most serious: the finger-ROM kinematic analysis is not valid. The paper says YOLO11 was fine-tuned on 11k Hands to obtain MCP/PIP/DIP angles, but 11k Hands is a detection/classification dataset, not a hand-pose keypoint dataset. There is no way to get joint angles from it, and there is no motion-capture validation. On top of that, the eye-in-hand and egocentric conditions mount the camera in physically different locations, so any ROM difference is confounded with viewpoint and possibly hand loading. Figure 4 cannot be attributed to the controller.\n\nSo my verdict: the integration and the GAS comparison are worth a serious look, but the paper as written is not self-consistent. The kinematic-alignment claim should be removed or redone with a proper pose estimator and controlled camera placement; the ablation table must be fixed; the abstract must be brought in line with Table 1. This deserves peer review, not desk rejection, because the system is real and the core experiments are re-runnable. I'd send it back for major revision. For a reading group, it's a useful cautionary example of checking tables against claims.","headline":"A real integration paper whose own tables and methods undercut its headline: the ablation is internally reversed, the GAS comparison loses on pinch, and the finger-ROM analysis is built on an invalid pose-estimation pipeline.","tokens_in":15561,"tokens_out":3717,"would_cite":false,"duration_ms":31891,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A soft hand exoskeleton can read a scene and listen for commands, achieving an 87% grasps-ability score without buttons or touch.","keywords":["Grasping Assistance","Wearable hand exoskeleton","Contextual Awareness","Open-Vocabulary Detection","Multimodal Intention Detection","Egocentric Vision","Soft robotics"],"falsifier":"Run the same grasping protocol with marker-based optical motion capture on the index and middle fingers under the three conditions—exoskeleton off, eye-in-hand exoskeleton on, and egocentric exoskeleton on. If the measured MCP, PIP, and DIP range-of-motion differences between camera views disappear or reverse, the paper's kinematic alignment claim would be refuted.","tokens_in":14772,"feed_emoji":"🦾","tokens_out":3351,"duration_ms":33806,"temperature":0.7,"pith_summary":"This paper proposes OVGrasp, a hierarchical control pipeline for a soft hand exoskeleton that decides when to grip and release by combining egocentric RGB-D vision, depth, and voice. The central aim is to show that a wearable assistive glove can be driven by open-vocabulary object detection and speech instead of push buttons or force contact, making it more generalizable and more natural for users with motor impairments. Ten healthy participants grasped 15 objects with pinch, spherical, and cylindrical grips, and the system reached a Grasping Ability Score of 87.00 ± 2.46%, outperforming push-button and force-sensing baselines while preserving finger-joint motion more faithfully, especially when the camera is head-mounted rather than worn under the hand. If correct, this points toward hands-free, adaptive grasp assistance that can handle objects unseen during training.","feed_headline":"Exoskeleton glove reads scenes and voice to trigger grasps at 87%","feed_subtitle":"Vision-language detection plus spoken release commands beat push-button and force-sensing controls in trials with ten users.","key_machinery":"The load-bearing mechanism is the three-level hierarchy: a YOLO-World vision-language detector that matches region embeddings to open-vocabulary text prompts; a middle-level decision-maker that registers depth with RGB, builds a Grasp-Point Graph whose nodes are object centers, selects the nearest node as the target, and maps voice keywords through the FunASR speech recognizer into symbolic commands G, R, and S; and a low-level PID controller with velocity feedback and indirect torque limiting that drives the cable-driven glove. The Grasp-Point Graph is what converts raw detections into a stable, multi-object decision about which object to grasp.","core_discovery":"The paper's central claim is that a hierarchical controller—high-level open-vocabulary detection, mid-level multimodal intent fusion, and low-level PID actuation—can make a cable-driven soft exoskeleton trigger grasps and releases in a natural, context-aware way. The high-level detector uses a vision-language model to recognize both pre-trained and previously unseen objects from plain text prompts. The middle-level decision-maker builds a grasp-point graph from detected objects and depth data, selects the nearest object as the target, and confirms grip commands while voice keywords trigger release and stop. Deployed on a custom egocentric-view exoskeleton, this arrangement scores higher on t","pith_inferences":["Inference: because the open-vocabulary detector is prompt-driven, the same architecture could be extended beyond grasping to other assistive hand actions—pointing, pinching, or tool use—by replacing the symbolic command set.","Inference: the speech-based release channel presumes intelligible vocal commands; for users with dysarthria or limited speech, the same mid-level decision layer could instead fuse eye gaze or residual muscle signals without changing the rest of the hierarchy.","Inference: the kinematic results are presented as biomechanical evidence, but the claim that egocentric views preserve natural motion would be stronger if the pose-estimation joint angles were validated against motion capture; that validation is not in the paper.","Inference: the nearest-object rule works well in single-target trials, but in denser scenes a more explicit disambiguation step—e.g., combining gaze or pointing with voice—would likely be needed before clinical deployment."],"forward_implications":["Users can add new object categories by simply typing a text prompt, so the glove can adapt to personal, open-ended environments without retraining.","Voice release and stop commands let the system handle the full grasp-release cycle, not just the initial grip trigger, which is a noted gap in vision-based assistive gloves.","An egocentric, head-mounted camera avoids the physical collision of an under-hand camera with the table, preserving more natural finger joint motion during low-profile and cylindrical grasps.","The multimodal intent fusion is designed for multi-object scenes, allowing the nearest object to be selected automatically while speech provides an extra safety valve.","Comparing the hierarchy against push-button and force-sensing controllers on the same glove suggests that context-aware visual triggering frees the user's attention for hand placement rather than for pressing or sensing contact."],"supporting_citations":[{"why":"Supplies the YOLO-World open-vocabulary detector architecture used as the high-level visual module.","marker":"[17]"},{"why":"Provides the FunASR speech recognition model that converts voice commands into release and stop signals.","marker":"[18]"},{"why":"Supplies the base cable-driven soft exoskeleton design and the push-button and force-sensing baselines compared in the study.","marker":"[22]"},{"why":"Defines the grasping and maintaining scoring protocol used to compute the Grasping Ability Score.","marker":"[44]"},{"why":"Defines the Anthropomorphic Hand Assessment Protocol and the Grasping Ability Score metric.","marker":"[45]"},{"why":"Supplies the YOLO11 model fine-tuned for hand pose estimation used in the finger joint ROM analysis.","marker":"[47]"},{"why":"Supplies the 11k Hands dataset used to fine-tune the pose model for finger joint angle measurement.","marker":"[48]"},{"why":"Supplies the YCB object set used for seen objects and the grasp-type categorization of the test objects.","marker":"[43]"},{"why":"Provides the prior egocentric-view intention-detection approach that motivates the head-mounted camera choice.","marker":"[10]"}],"fun_headline_variants":["Open-vocab exoskeleton grasps unseen objects via voice and vision","Multimodal glove: voice and vision drive 87% grasp success on new objects","Glove reads words and scenes to grasp any object without retraining","87% grasp assist from voice+vision exoskeleton beats push-button controls","Zero-shot grasping exoskeleton uses text prompts and voice to aid users"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The finger-joint motion results assume the pose-tracking model measures joint angles accurately and treats both camera views equally, but no motion-capture validation was done and the two exoskeleton-on conditions use different camera mountings.","fun_headline_variants_meta":{"raw":{"variants":["Open-vocab exoskeleton grasps unseen objects via voice and vision","Multimodal glove: voice and vision drive 87% grasp success on new objects","Glove reads words and scenes to grasp any object without retraining","87% grasp assist from voice+vision exoskeleton beats push-button controls","Zero-shot grasping exoskeleton uses text prompts and voice to aid users"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1250,"prompt_tokens":703,"completion_tokens":547,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":447,"completion_tokens_details":{"reasoning_tokens":461}},"tokens_in":447,"tokens_out":547,"duration_ms":5192,"temperature":1.0,"reasoning_tokens":461,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:13:03.399869+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same grasping protocol with marker-based optical motion capture on the index and middle fingers under the three conditions—exoskeleton off, eye-in-hand exoskeleton on, and egocentric exoskeleton on. If the measured MCP, PIP, and DIP range-of-motion differences between camera views disappear or reverse, the paper's kinematic alignment claim would be refuted.","supporting_citations":[{"cited_title":"Cheng, L","cited_arxiv_id":null,"evidence_quote":"Supplies the YOLO-World open-vocabulary detector architecture used as the high-level visual module."},{"cited_title":"Point Cloud-based Grasping for Soft Hand Exoskeleton","cited_arxiv_id":"2504.03369","evidence_quote":"Supplies the base cable-driven soft exoskeleton design and the push-button and force-sensing baselines compared in the study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the grasping and maintaining scoring protocol used to compute the Grasping Ability Score."},{"cited_title":"Llop-Harillo, et al., The anthropomorphic hand assessment protocol (ahap), Robotics and Autonomous Systems 121 (2019) 103259","cited_arxiv_id":null,"evidence_quote":"Defines the Anthropomorphic Hand Assessment Protocol and the Grasping Ability Score metric."},{"cited_title":"Jocher, J","cited_arxiv_id":null,"evidence_quote":"Supplies the YOLO11 model fine-tuned for hand pose estimation used in the finger joint ROM analysis."},{"cited_title":"Calli, Benchmarking in manipulation research: Using the yale-cmu- berkeley object and model set, IEEE Robotics & Automation Magazine 22 (3) (2015) 36–52","cited_arxiv_id":null,"evidence_quote":"Supplies the YCB object set used for seen objects and the grasp-type categorization of the test objects."},{"cited_title":"Kim, et al., Eyes are faster than hands: A soft wearable robot learns user intention from the egocentric view, Science Robotics 4 (26) (2019) eaav2949","cited_arxiv_id":null,"evidence_quote":"Provides the prior egocentric-view intention-detection approach that motivates the head-mounted camera choice."}],"review_version":1}