{"id":"231ad191-b599-4c46-9a24-40b7bcfd60c3","arxiv_id":"2505.01974","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"KineDex shows that kinesthetic teaching with inpainting and force-informed actions trains tactile-aware visuomotor policies that outperform position-only control on nine dexterous manipulation tasks.","lead":"This paper introduces KineDex, a hand-over-hand kinesthetic teaching system that records tactile-enriched demonstrations by having a person physically guide a robot hand. Policies trained on these demonstrations with force control succeed at 74.4% across nine contact-rich tasks, and data collection is more than twice as fast as teleoperation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Force-control law in Eq. (2) is an open-loop, fixed-gain position offset that ignores object stiffness; the reported ablation gains may not be due to accurate force tracking.","rationale":"The reader identified the fixed-gain force-control law as the weakest assumption, and I agree: the central advantage claimed for KineDex over the no-force-control variant rests on the mechanism in Eq. (2). My reading adds two concrete technical details that make the concern more specific. First, the law is open-loop with respect to force: there is no feedback from measured force to the commanded offset, so the actual force is set by the environment's stiffness rather than by f_d. Since the tasks use objects with very different contact stiffnesses, the fixed-gain assumption is not just a tuning concern but a structural limitation of the control law. Second, the equation as written uses the current position rather than the policy's predicted position x_d, which makes the policy's joint-position output irrelevant to the controller; if f_d = 0, the hand is servo-locked. This suggests either a typo in Eq. (2) or an under-specified control architecture. The paper does provide real empirical evidence in the form of a 20-trial-per-task ablation table, and the inpainting and tactile ablations are informative, so I would not reject the paper based on this concern alone. However, because the force-control mechanism is the stated source of the largest performance gain, the manuscript needs either a corrected control-law derivation or a direct validation of force tracking across objects of varying stiffness before the central claim can be accepted as described. The reader's conditional verdict remains appropriate, and my read does not change it.","tokens_in":12094,"tokens_out":6777,"duration_ms":80651,"concrete_test":"Re-run inference on two tasks with very different object compliance, e.g., Egg Picking and Syringe Pressing, using the same fixed K_tip and K_base, while logging desired f_d and measured fingertip force. If the measured-to-desired force ratio differs by more than about 30% between tasks, the fixed-gain linear stiffness assumption in Eq. (2) fails and the stated force-control mechanism cannot explain the reported gains. Additionally, implement Eq. (2) literally with the policy's x_d removed from the hand PD target; check whether any non-grasping motion such as cap twisting is possible when f_d = 0. If the hand is servo-locked at f_d = 0, the controller description is internally inconsistent and must be corrected before the central force-control claim can be evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The force-control mechanism in Eq. (2) is not a force controller in the usual sense. It commands x_tip^d = x_tip + K_tip * f_d and x_base^d = x_base + K_base * f_d using the current measured positions, with no feedback term comparing measured force to f_d; tactile signals enter only the policy input, not the control loop. The gains K_tip and K_base are tuned once and fixed across all tasks, so for an object of stiffness k the steady-state contact force is approximately k * K * f_d plus preload, not f_d itself. The nine tasks include raw egg, paper cup, plastic bottle, and syringe, whose contact stiffnesses differ by orders of magnitude, so the same fixed gains cannot keep actual force close to the predicted force across tasks. The large improvement over the no-force-control variant (0/20 versus 15/20 on several tasks) may therefore reflect the effect of an additive preload offset rather than accurate force tracking. Figure 3 shows tracking on one task only and does not validate the fixed-gain assumption across tasks. In addition, Eq. (2) as written never uses the predicted joint positions x_d: if f_d = 0, the target equals the current position, so the hand would be servo-locked and the role of x_d in the action space (Section 3.1) is vacated. Either the equation is missing an x_d term or the reported executions use a different, unspecified control law; neither possibility is compatible with the claim that KineDex 'precisely track[s] the target fingertip forces.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"KineDex proposes a kinesthetic-teaching pipeline for collecting demonstrations with a dexterous hand, with tactile sensing and fingertip normal forces recorded. The raw visual observations are preprocessed by segmenting and inpainting the operator's body, and a Diffusion Policy is trained to predict both target joint positions and target fingertip normal forces from visual, tactile, and proprioceptive inputs. At deployment, the predicted forces are used to modify desired position targets through a proposed force-control law in Eq. (2). The paper evaluates the method on nine contact-rich manipulation tasks, reporting an average success rate of 74.4% for KineDex versus 16.7% without force control, and also reports a data-collection comparison against a teleoperation baseline and a small user study. The central claim is that kinesthetic teaching provides efficient, tactile-enriched demonstrations that, combined with tactile-informed policies and force control, produce reliable policies for contact-rich manipulation.","tokens_in":12347,"tokens_out":4222,"duration_ms":43455,"significance":"If the claims hold, the paper makes a useful empirical contribution: it demonstrates a practical kinesthetic-teaching setup for a dexterous hand, shows that inpainting occluded operators is a workable alternative to trajectory replay, and provides evidence that tactile-augmented actions improve contact-rich manipulation. The data-collection efficiency comparison against teleoperation is valuable and, if reproduced, would be a practical advantage. However, the load-bearing force-control contribution is currently underspecified and the quantitative evidence is limited to 20 trials per condition without variance reporting, so the significance can be assessed fully only after the control law and experimental statistics are clarified.","major_comments":[{"comment":"Equation (2) is not a force controller in the usual sense, and as written it does not use the predicted joint positions x_d. The force-informed target positions are computed from the current measured positions plus a fixed gain times the predicted force, with no term comparing measured force to the predicted force and no mention of x_d. If f_d = 0, the target equals the current position, so the position-control loop would be servo-locked and the joint-position component of the action would be vacated. Either the equation is missing an x_d-dependent term or the deployed controller differs from the stated one. The claim that KineDex \"precisely tracks\" the target forces is therefore unsupported by the given control law, and the implementation must be specified exactly.","section":"§3.4, Eq. (2)"},{"comment":"The force mechanism relies on fixed gains K_tip and K_base that are described as tuned once and kept fixed across tasks. Because the nine tasks involve objects with very different contact stiffnesses (raw egg, paper cup, plastic bottle, syringe), an open-loop position offset proportional to f_d cannot be expected to produce the same actual force across tasks; the steady-state contact force is approximately stiffness times the offset, not f_d itself. The paper should either provide a closed-loop force controller, report measured force-tracking errors per task, or restrict the claim to demonstrating that an additive preload offset helps rather than that the policy accurately tracks predicted forces.","section":"§3.4, Eq. (2)"},{"comment":"All performance claims rest on a single set of 20 trials per condition, with no repeated seeds, no standard deviations, and no confidence intervals. With 20 binary trials, the standard error of a 75% success rate is roughly 10 percentage points, so differences such as 12/20 versus 16/20 (Charger Plugging between KineDex and w/o Tactile Input) are not statistically meaningful. Reporting multiple training seeds and trial-level variation is necessary to support the headline 74.4% average and the 57.7% improvement claim.","section":"Table 1, §4.1"},{"comment":"The w/o Inpainting ablation yields zero successes on all nine tasks, including simple pick-and-place tasks such as Bottle Picking and Cup Picking. This is a surprising and strong result that is not supported by any analysis of the failure modes, the quality of the inpainted images, or the sensitivity of the result to inpainting artifacts. Given that this ablation is used to justify the inpainting component of the pipeline, the paper should provide more than a single zero-success observation, for example by reporting policy behavior on raw versus inpainted observations or by varying the inpainting quality.","section":"Table 1, §4.1"}],"minor_comments":[{"comment":"The phrase \"tactile-enriched demonstrations\" is followed by \"viatactile-informed\" in the Figure 1 caption, which appears to be a missing-space typo.","section":"Abstract and Figure 1 caption"},{"comment":"The notation \"No\" in the observation space description appears to be a corrupted subscript; if it denotes N_o views, the text should say so explicitly.","section":"§3.1"},{"comment":"The action vector is described as 23-dimensional (6 end-effector DoF, 12 joint angles, 5 force targets), but §3.1 defines the action as only target joint positions and target forces, omitting the end-effector pose. These two descriptions should be reconciled.","section":"§B, Policy Training Details"},{"comment":"The paper states that the operator controls the thumb with the left hand while guiding the other fingers with the right hand, but Figure 4 and the task descriptions do not clarify how this bimanual coordination is synchronized or whether it affects the quality of force demonstrations. A brief explanation would help reproducibility.","section":"§3.2"},{"comment":"The force-tracking visualization uses normalized forces and shows only two tasks; the paper should state the scaling and, ideally, report quantitative tracking error or correlation between predicted and sensed force.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The empirical direction is promising and the kinesthetic-teaching comparison is likely of interest to the robotics community. My main concern is that the force-control method, which is central to the paper's claims, is not described precisely enough to reproduce and appears to be an open-loop position offset rather than a force controller. I would be willing to evaluate a revised version that specifies the exact control law, adds force-tracking measurements across tasks, and reports variances or multiple seeds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nKineDex is a genuine step forward in collecting tactile-rich demonstrations for dexterous hands. The core idea—hand-over-hand kinesthetic teaching with visual inpainting to remove the operator, then training a diffusion policy on tactile-plus-visual observations—is clearly presented and the empirical setup is substantial. Nine contact-rich tasks, including raw egg picking and toothpaste squeezing, with 20 trials each, is more than most papers in this space. The data collection efficiency over teleoperation is credible and supported by a user study. The ablation without inpainting giving zero across all tasks shows why the preprocessing matters, and the force-control variant's improvement over position-only control is large.\n\nThe soft spot is the force-control law. Equation (2) defines the force-informed target position as x_tip + K_tip * f_d (and analogously for the base), with no dependence on the predicted joint positions x_d. If f_d is zero, the target equals the current position, which means the hand would be servo-locked and the role of x_d in the action space is vacated. Either the equation is missing an x_d term, or the actual execution uses a different, unspecified control law. That matters because the paper claims to 'precisely track' the target fingertip forces, and the only supporting figure shows tracking on one task. The gains K_tip and K_base are tuned once, not reported, and the steady-state force depends on object stiffness; for a raw egg, a paper cup, and a syringe, the same fixed gains cannot keep actual force close to the predicted force. The improvement over the no-force-control variant might come from an additive preload rather than accurate force tracking. The authors need to clarify this.\n\nAlso, the experimental reporting is thin: no repeated seeds, no standard deviations. Twenty trials per cell gives some signal, but for a paper whose headline number is a 57.7% improvement (which is actually percentage points, not a relative improvement), the lack of error bars is noticeable. No code or data is released, which makes it harder to verify the surprising zero success without inpainting.\n\nWho is this for: anyone working on imitation learning for dexterous manipulation, especially on data collection. It deserves a serious referee; the issues above are fixable with an explicit description of the control law and a few more experimental details. I would not cite it in its current form, but I'd accept it for peer review and ask for a revision.","headline":"Useful kinesthetic-teaching framework with a real contribution in visual inpainting, but the force-control law in Eq. (2) is under-specified and undermines the central claim; needs revision before I'd trust the numbers.","tokens_in":12900,"tokens_out":4651,"would_cite":false,"duration_ms":41257,"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":"KineDex claims that kinesthetic teaching—letting a human physically guide the robot hand during demonstrations—combined with inpainting the operator out of the view and adding force control at deployment, trains tactile-informed…","keywords":["dexterous manipulation","kinesthetic teaching","tactile sensing","force control","visuomotor policy","imitation learning","image inpainting","contact-rich manipulation"],"falsifier":"Run the trained KineDex policy on a set of objects matched to the nine tasks but with substantially different surface stiffness, such as a rigid steel bottle, a hard plastic cap, and a soft foam syringe, without retuning $K_{\\mathrm{tip}}$ and $K_{\\mathrm{base}}$. If the executed fingertip forces, measured by the hand's tactile sensors, systematically deviate from the policy's predicted forces, or if success rates fall to the no-force-control level on the rigid object, the linear fixed-gain force law is the failing component. A weaker test is to record force tracking on all five fingers, not just the thumb, and check that predicted and sensed forces match at contact.","tokens_in":11868,"feed_emoji":"🖐️","tokens_out":7872,"duration_ms":69227,"temperature":0.7,"pith_summary":"KineDex claims that letting a human physically guide a robot hand during demonstrations rather than teleoperating it from a distance produces the high-fidelity tactile data needed to learn contact-rich manipulation, and that this data, once the operator's body is inpainted out of the camera views, can train a visuomotor policy that succeeds on objects as delicate as raw eggs and as force-demanding as syringes. The paper reports an average success rate of 74.4% across nine tasks, a 57.7% improvement over the same policy without force control, and shows the no-force-control variant fails almost completely on tasks like peg insertion and charger plugging. A careful reader would care because it offers a concrete data-collection alternative to teleoperation that is roughly twice as fast and far more reliable during collection, while addressing the occlusion problem that previously forced trajectory replay. The load-bearing claim is that force feedback during both teaching and execution is what unlocks these tasks, not just better vision.","feed_headline":"Physical teaching lifts dexterous-robot success to 74.4 percent","feed_subtitle":"Direct force feedback and inpainted views beat position-only control by 57.7 percent on nine contact-rich tasks.","key_machinery":"The central mechanism is the force-informed target position, a control law that converts each fingertip's predicted contact force into a virtual displacement of the fingertip and base joints before the PD controller runs. For each finger the target positions become $x_{\\mathrm{tip}}^d = x_{\\mathrm{tip}} + K_{\\mathrm{tip}} f^d$ and $x_{\\mathrm{base}}^d = x_{\\mathrm{base}} + K_{\\mathrm{base}} f^d$, where $f^d$ is the normal force predicted by the policy and $K_{\\mathrm{tip}}, K_{\\mathrm{base}}$ are fixed gains; pressing the virtual target inside the object generates real contact force through the object's resistance. This is what allows the trained policy to track the demonstration's force profile instead of merely touching the surface. The other load-bearing pieces are the inpainting preprocessing—a segmentation step plus a video inpainting model that removes the operator's body from the front camera—and the tactile encoder that feeds per-finger contact vectors into the policy alongside images and proprioception. The force control law is the primary carrier of the paper's success gains.","core_discovery":"The paper's central claim is that kinesthetic teaching with a hand-over-hand setup, combined with image inpainting and tactile-augmented policy learning, is sufficient to train visuomotor policies that succeed on nine contact-rich dexterous manipulation tasks. The key demonstration is that when the learned policy predicts both target joint positions and target fingertip forces, and those forces are converted into virtual joint offsets during execution, KineDex reaches an average success rate of 74.4%, with about 15 of 20 trials or better on most tasks; removing force control collapses performance to 16.7% average, with zero successes on Bottle Picking, Peg Insertion, Charger Plugging, Toothpaste Squeezing, and Syringe Pressing. The paper also shows that tactile input adds a 26.7% average improvement on contact-intensive tasks such as Cap Twisting, Toothpaste Squeezing, and Syringe Pressing, and that inpainting the operator's hand out of demonstrations is necessary because without it success is zero across all tasks. Comparatively, KineDex collects demonstrations more than twice as fast as the teleoperation baseline while maintaining near-100% collection success, versus under 50% for teleoperation. The intended conclusion is that physically grounded tactile demonstrations plus force-informed control, not vision alone, are what make precise contact-rich manipulation learnable.","pith_inferences":["A natural extension is to make the force-control gains adaptive, for instance by estimating surface stiffness from tactile readings during contact, so the same policy transfers to objects with very different compliance without retuning.","The inpainting step is a likely bottleneck for scenes with severe or dynamic occlusion; a testable improvement is to train the inpainting model on robot-task data, which the paper itself flags as future work.","The reliance on two human hands, one for the thumb and one for the other fingers, suggests the method will not scale directly to bimanual tasks until more biomimetic hardware or single-handed guidance is available; meanwhile, the comparison with teleoperation is the key evidence that the kinesthetic channel, not just extra demonstration count, drives the gain.","One could test whether the force prediction itself can be learned from teleoperation or video demonstrations paired with a force-sensor-equipped hand, which would decouple the kinesthetic collection advantage from the force-control advantage."],"forward_implications":["If KineDex is right, kinesthetic teaching can replace teleoperation as the default way to collect dexterous-hand demonstrations: it is roughly twice as fast, needs no VR retargeting, and gives operators natural force feedback.","The force-informed position offset makes the gap between touching an object and applying the right force learnable; position-only policies will keep failing on contact-rich tasks even with good vision.","Inpainting the operator out of the view turns kinesthetic demos into usable training data, eliminating the need to replay trajectories for clean observations, so longer-horizon tasks become feasible.","Tactile sensing is not just auxiliary: on tasks where vision is occluded or contact feedback is central, such as cap twisting, toothpaste squeezing, and syringe pressing, removing it costs an average 26.7% success.","Success on deformable and fragile objects such as eggs, paper cups, and syringes suggests the same pipeline may extend to other precision tasks in assembly, healthcare, or service robots."],"supporting_citations":[{"why":"The closest prior work that also collects tactile-augmented demonstrations via kinesthetic teaching; KineDex extends it to higher-DOF hardware and longer-horizon tasks.","marker":"[28]"},{"why":"Supplies the diffusion-based visuomotor policy learning method that KineDex uses as its backbone and augments with tactile input and force-informed actions.","marker":"[44]"},{"why":"The teleoperation system that KineDex replicates as its baseline for data-collection efficiency and success-rate comparison.","marker":"[15]"},{"why":"Provides the video inpainting model that removes the operator's body from the visual demonstrations, making the data usable for policy training.","marker":"[30]"},{"why":"The segmentation model used to generate masks of the operator's body parts for the inpainting preprocessing.","marker":"[51]"},{"why":"Provides the action-chunking strategy that smooths the policy's control during inference, improving temporal consistency.","marker":"[52]"},{"why":"Used for retargeting human hand keypoints to the robot hand in the teleoperation baseline, enabling the fair comparison of data collection methods.","marker":"[31]"}],"fun_headline_variants":["Tactile-informed policy lifts robot dexterity to 74.4% success","Kinesthetic teaching with force feedback: 74.4% dexterous success","Force-aware visuomotor policy: 57.7-point gain over position-only control","Hand-over-hand kinesthetic teaching: 74.4% success on nine tasks","Touch-guided teaching achieves 74.4% success on contact-rich tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gains $K_{\\mathrm{tip}}$ and $K_{\\mathrm{base}}$ in the force-to-position law are tuned once and fixed across every task and object, so the claim that force control dramatically improves success rests on the assumption that a single linear force-displacement relation holds for all the objects the hand touches.","fun_headline_variants_meta":{"raw":{"variants":["Tactile-informed policy lifts robot dexterity to 74.4% success","Kinesthetic teaching with force feedback: 74.4% dexterous success","Force-aware visuomotor policy: 57.7-point gain over position-only control","Hand-over-hand kinesthetic teaching: 74.4% success on nine tasks","Touch-guided teaching achieves 74.4% success on contact-rich tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001281,"raw_usage":{"total_tokens":5323,"prompt_tokens":1123,"completion_tokens":4200,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":739,"completion_tokens_details":{"reasoning_tokens":4094}},"tokens_in":739,"tokens_out":4200,"duration_ms":29933,"temperature":1.0,"reasoning_tokens":4094,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:03:54.184042+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained KineDex policy on a set of objects matched to the nine tasks but with substantially different surface stiffness, such as a rigid steel bottle, a hard plastic cap, and a soft foam syringe, without retuning $K_{\\mathrm{tip}}$ and $K_{\\mathrm{base}}$. If the executed fingertip forces, measured by the hand's tactile sensors, systematically deviate from the policy's predicted forces, or if success rates fall to the no-force-control level on the rigid object, the linear fixed-gain force law is the failing component. A weaker test is to record force tracking on all five fingers, not just the thumb, and check that predicted and sensed forces match at contact.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The closest prior work that also collects tactile-augmented demonstrations via kinesthetic teaching; KineDex extends it to higher-DOF hardware and longer-horizon tasks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the diffusion-based visuomotor policy learning method that KineDex uses as its backbone and augments with tactile input and force-informed actions."},{"cited_title":"Cheng, J","cited_arxiv_id":null,"evidence_quote":"The teleoperation system that KineDex replicates as its baseline for data-collection efficiency and success-rate comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the video inpainting model that removes the operator's body from the visual demonstrations, making the data usable for policy training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The segmentation model used to generate masks of the operator's body parts for the inpainting preprocessing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the action-chunking strategy that smooths the policy's control during inference, improving temporal consistency."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Used for retargeting human hand keypoints to the robot hand in the teleoperation baseline, enabling the fair comparison of data collection methods."}],"review_version":1}