{"id":"cca5ae6a-ce3f-4203-9165-edfa071fb684","arxiv_id":"2507.07980","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A neural network trained on real joint torque and position data localizes touches on a robot's body, reaching 7.2 cm average error on Spot and 8.0 cm on Franka without tactile sensors.","lead":"UniTac uses only the joint sensors already inside a robot to guess where it is being touched, with no extra tactile skin. On a Spot quadruped and a Franka arm it localizes contacts to about 7 to 8 centimeters at roughly 2,000 Hz, which could make basic touch sensing available on ordinary robots.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported errors are measured only at the discrete training points, so the central claim of whole-robot, arbitrary-contact localization is not yet established; leave-one-point-out evaluation would test it directly.","rationale":"I read the paper in good faith. The method is plausible: static joint torques under varied poses carry enough information to localize contact, and the data collection effort is non-trivial. The architecture is simple but adequate for the task. The paper also provides qualitative real-time deployment and a qualitative cross-instance generalization result, which count as modest supporting evidence. However, the central quantitative claim is not yet about arbitrary contact: all ground-truth labels come from a small finite set, and test samples are drawn from that same set under the same protocol. This is exactly the reader's weakest assumption, with which I agree. The proposed leave-one-point-out test would settle whether the learned mapping generalizes continuously beyond the training locations, or whether the 7-to-8 cm error is a memorization artifact. It is a concrete, low-cost experiment using existing data. Because the reader already rendered a CONDITIONAL verdict on this basis, I recommend no change in verdict. I do not see a basis for rejection: there is no evidence of internal inconsistency or fabrication, and the qualitative demonstrations partially support the claim. The concern is about scope of evidence, not soundness of the method.","tokens_in":10176,"tokens_out":5597,"duration_ms":66269,"concrete_test":"Re-run the Spot evaluation with leave-one-point-out: train on 103 of the 104 contact points, test on the held-out point, and repeat for all 104 points (same for Franka's 10 points). If the average L2 error on held-out points remains near the reported 7.2 cm / 8.0 cm, the model is interpolating across the body surface and the central claim is supported. If the error rises substantially (e.g., above 15 cm), the reported accuracy is an artifact of testing on trained discrete locations and the whole-robot claim would need to be scaled back to trained-location recognition only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is that the headline numbers only measure localization at the discrete, hand-selected contact points used as training labels. In Sec. 4.1.1-4.1.2, Spot is touched at 104 fixed mesh points (legs excluded) and Franka at 10 fixed points, and each ground truth p_i in the dataset is one of these points. The regression model (Sec. 3.1) is trained with MSE against these coordinates, and the 80/20 split in Sec. 4.2.1 simply reserves 20% of samples from the same point set for evaluation. The reported 7.2 cm / 8.0 cm L2 therefore answers 'given a torque signature from one of the 104 (or 10) known body locations, how close is the prediction to that known location?' It does not answer 'given a touch at an arbitrary, previously unseen body location, where is it?' An MLP that has effectively memorized the discrete location set can score well on this protocol; the KNN baselines in Table 1 are likewise evaluated on the same discrete set and do not bound continuous generalization. Since the title and abstract claim whole-robot touch sensing, and the applications (Sec. 5) require responding to natural touches anywhere on the body, the missing continuous and interpolation evaluation is the load-bearing gap. The qualitative sliding-touch results (Fig. 4-5) are suggestive but not quantitative, and the cross-instance Spot demonstration is also qualitative. This is not an internal inconsistency in the method; it is an external-validity gap in the evidence for the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UniTac, a supervised method that maps joint position and torque readings to 3D contact coordinates using a four-layer MLP, without any tactile hardware. Training data are collected by human touches at 104 sampled points on a Spot quadruped (with the legs excluded) and 10 points on a Franka arm, under randomized joint configurations. On an 80/20 split of samples drawn from these discrete point sets, UniTac reports average L2 localization errors of 7.2 cm on Spot and 8.0 cm on Franka, at roughly 2,000 Hz on an RTX 3090 GPU. The paper also presents qualitative pHRI demonstrations on both platforms and a qualitative cross-instance generalization test on a second Spot robot.","tokens_in":10470,"tokens_out":4721,"duration_ms":51455,"significance":"If the reported localization accuracy held for arbitrary, previously unseen contact locations, UniTac would be a practical and useful contribution to physical human-robot interaction, because it requires no additional sensors and runs at high frequency. The paper's strengths are its real-world data collection on two distinct embodiments, its simple and reproducible network design, and its concrete demonstrations on a quadruped and an arm. However, the current evaluation protocol draws test samples from the same finite set of discrete contact points used as training labels, so the headline errors measure recognition of known locations rather than continuous whole-body localization. The 'whole-robot' claim is also weaker than stated because Spot's legs are excluded from data collection. These gaps are fixable with additional experiments and do not undermine the core supervised-learning formulation itself.","major_comments":[{"comment":"The quantitative evaluation only holds out samples from the same discrete set of contact points used as training labels (104 on Spot, 10 on Franka; §4.1.1–4.1.2). The reported L2 errors therefore measure how well the model identifies one of the known training locations, not how well it localizes a contact at an arbitrary, previously unseen body location. Because the regression output is continuous but all ground-truth labels come from the finite sampled set, a model that effectively memorizes the set could still score well on this protocol. I request leave-one-point-out evaluation and/or a test set of touches at locations not in the training set, with continuous ground truth (e.g., motion capture or a tracked probe), and with per-point errors reported. This is the central missing evidence for the 'whole-robot' claim.","section":"§4.2.1, Table 1"},{"comment":"The paper uses 'whole-robot' and 'whole-body' in the title, abstract, and introduction, but on Spot the legs are explicitly excluded from data collection and evaluation: the Figure 3 caption states that the 104 sampled points 'cover the whole robot except for the legs.' The claim of whole-robot touch sensing therefore overstates the demonstrated coverage. Please either qualify the claims to the covered body regions or extend data collection and evaluation to the legs and any other uncovered surfaces.","section":"§4.1.1, Figure 3"},{"comment":"Accuracy is reported at a single threshold epsilon = 12 cm without justification, and all quantitative results are point estimates from one 80/20 split, with no error bars or multiple seeds. Since the central claims rest on these numbers, the paper should report mean and standard deviation over several training seeds and data splits, and should justify the threshold or report threshold-curve statistics (e.g., area under the accuracy-versus-epsilon curve in Figure 6). Without this, the Acc values in Table 1 are not statistically interpretable and the comparison between methods is fragile.","section":"§4.2.1, Eq. (1)"},{"comment":"The claimed generalization to a different Spot instance is supported only by qualitative results: the text says that sliding-touch localization on an unseen Spot 'exhibit[s] similar accuracies' but no quantitative numbers are given. Given the stated goal of an off-the-shelf interface for other robots, please provide a quantitative evaluation on the second Spot instance using the same protocol and metrics as Table 1, and ideally also with contact locations not used during training. Without such data, the cross-instance generalization claim is not supported.","section":"§4.2.2, Figure 4"}],"minor_comments":[{"comment":"The regression model represents the 'no-contact' state by the point (0,0,0), but the data collection and evaluation sections do not clearly describe how no-contact samples are collected and labeled, or whether the reported metrics include them. Clarify the role of no-contact data in training and in the quantitative results.","section":"§3.1, §4.1"},{"comment":"Please label the axes and specify whether the accuracy-threshold curves correspond to Spot, Franka, or both. The caption currently says only that accuracy is evaluated at 'various threshold values.'","section":"Figure 6"},{"comment":"The statement that 'the robot is kept stationary for each sample' is important for understanding the evaluation scope, but it appears only inside the data collection paragraph. State this assumption explicitly in the method overview and in the limitations discussion, since the applications imply the robot may be touched during motion.","section":"§4.1.1"},{"comment":"The sliding-touch qualitative results are promising, but a quantitative trajectory error (for example, comparing the predicted contact path with the actual sliding path over time) would substantially strengthen the claim that the method tracks continuous, moving contacts.","section":"§4.2.2"},{"comment":"The virtual-button demonstration uses stickers at three locations on Franka; please clarify whether these locations coincide with the 10 sampled training points, and if so, whether the demonstration therefore only tests recognition of known locations rather than arbitrary button placement.","section":"§5.2.1"}],"recommendation":"major_revision","confidential_remarks":"I found no circularity in the core supervised-learning result; the main issue is external validity of the evaluation rather than internal inconsistency. The paper is a reasonable systems contribution, but the headline 'whole-robot, arbitrary contact localization' claim needs the additional experiments described in the major comments before it can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. UniTac is a genuinely practical result: real-world data, two embodiments, no tactile hardware, and a claimed cross-instance transfer. The 7.2 cm / 8.0 cm errors at 2 kHz are plausible for the task as defined. The second thing is that the evaluation only tests touches at the discrete, hand-selected points used for training labels. The stress-test note lands: the 80/20 split draws held-out samples from the same point set, so the numbers do not demonstrate arbitrary-contact localization anywhere on the body.\n\nWhat is new: prior data-driven contact localization used simulation and single arms; this paper trains on real torque/position data for a quadruped and an arm, and shows qualitative transfer to an unseen Spot. That is a step forward. The data collection effort is modest and described clearly, and the regression-vs-classification comparison is a useful ablation.\n\nSoft spots. The main one is external validity: the regression labels are the coordinates of the sampled points, and evaluation is on samples from those points. A leave-one-point-out test, or touches at points not in the training set, would speak directly to the whole-robot claim. The 'whole-robot' wording is also generous: Spot legs are excluded. The paper does not compare against model-based torque-only methods such as Iskandar et al., which is the most relevant baseline. There are no error bars or multiple seeds, and no code or data, so the reported numbers cannot be independently checked. The epsilon=12 threshold is arbitrary but the L2 metric is there; that is a minor issue. The paper does honestly state it is limited to single-contact localization.\n\nThe central result is not broken. If you accept the discrete-point framing, a supervised model can learn a torque-to-contact map and interpolate reasonably as the qualitative sliding-touch demos suggest. The paper just sells it as more than it has shown.\n\nWho this is for: robotics and HRI researchers who want a no-hardware touch sensing option and are comfortable with the accuracy/resolution tradeoff. It deserves a serious referee. The revision should add the continuous/evaluation experiments and the missing baseline. If those come back, this becomes a solid applied contribution.\n\nRecommendation: accept for peer review, with major revision expected.","headline":"Useful no-hardware contact localization on real robots, but the headline errors only cover a discrete set of training points, so the whole-robot claim needs a continuity test.","tokens_in":11050,"tokens_out":2678,"would_cite":false,"duration_ms":29870,"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":"Using only the joint torque and position sensors already on a robot, UniTac learns to localize touches, reaching average errors of 7.2 cm on a Spot quadruped and 8.0 cm on a Franka arm at about 2,000 Hz.","keywords":["tactile sensing","contact localization","proprioception","joint torque sensing","human-robot interaction","quadruped robot","robot arm","data-driven robotics"],"falsifier":"Take a Spot or Franka, train UniTac on the paper's sampled contact points, then touch a location between those points (or on a Spot leg, which the paper excluded) while the robot is stationary, with ground truth measured by a motion-capture marker or a reference tactile patch. If the average localization error on these continuous, out-of-distribution touches is substantially larger than the reported 7-8 cm, then the 'whole-body' claim holds only for the discrete, hand-picked contacts the model was trained on.","tokens_in":9947,"feed_emoji":"🖐️","tokens_out":7268,"duration_ms":77608,"temperature":0.7,"pith_summary":"The paper tries to establish that whole-body touch sensing is possible on ordinary robots that lack tactile skins, using only the joint torque and position sensors already built into commercial platforms. If true, contact localization becomes an off-the-shelf capability: a few hours of labeling where human touches land, and a robot can react to touch in real time without new hardware. The key result is a regression model that maps proprioceptive signals to a 3D contact point, with average errors of 7.2 cm on a Spot quadruped and 8.0 cm on a Franka arm at about 2,000 Hz. The authors also show the model transfers to a second Spot instance without retraining and powers touch-based interaction demonstrations.","feed_headline":"No skin needed: joint sensors localize touch to 7.2 cm","feed_subtitle":"UniTac maps torque and position to contact points on Spot and Franka, enabling touch-based robot interaction.","key_machinery":"The load-bearing object is UniTac-Net, a four-layer MLP with layer sizes 64, 128, 256, and 128, ReLU hidden activations, dropout 0.3, and a three-output regression head. Its input is the normalized proprioceptive vector of joint positions $q$ and torques $\\tau$ (19 DoF on Spot, 7 on Franka), and its output is a predicted contact point $p$ in the robot frame; the no-contact state is coded as $(0,0,0)$, and training uses mean squared error with the Adam optimizer at a fixed learning rate of $2.5 \\times 10^{-3}$ for 30 epochs. The mechanism works because different contact locations produce distinguishable joint-torque signatures across randomized poses, letting the network learn the mapping from sensor feedback to body coordinates. An exponential moving average filter with smoothing factor 0.1 and sliding window 40 turns the 2,000 Hz raw predictions into temporally stable live estimates.","core_discovery":"UniTac's central claim is that whole-body touch localization can be learned from proprioception alone: a four-layer MLP maps joint positions and torques to 3D contact coordinates in the robot frame, reaching average L2 errors of 7.2 cm on a Spot quadruped and 8.0 cm on a Franka Research 3 arm at about 2,000 Hz on an RTX 3090 GPU. The authors treat this as a regression problem rather than classification, represent 'no contact' by predicting the origin, and stabilize live outputs with an exponential moving average filter. They further report that the model transfers to a second Spot instance without retraining and demonstrate real-time applications: touch regions on Spot trigger animal-inspired behaviors, and touch points on Franka act as virtual buttons that select which colored block the arm picks up.","pith_inferences":["A natural next test the paper does not run is continuous touch: sliding a finger along an arbitrary path between the sampled points and measuring error with an external tracker; the regression output suggests the model may interpolate, but the reported numbers do not establish it.","Because training needs only real-world joint data and a few hours of labeling, the same recipe should transfer to other torque-sensing morphologies such as bipeds or mobile manipulators, and to dynamic contacts during motion, both of which are outside the paper's stationary evaluation.","The paper's single-contact assumption leaves open the multi-contact case; a detection head or contact-count prior on top of the same proprioceptive features would be the natural extension for collision recovery and safe physical human-robot interaction."],"forward_implications":["A robot with no tactile skin can localize a touch on its body to about 7.2 cm on Spot and 8.0 cm on Franka using only joint torque and position data.","Regression beats classification by roughly 30 percentage points in accuracy and about 6.5 cm in L2 error, so predicting continuous coordinates is the better design under sensor noise.","The model localizes touches on a second, unseen Spot instance without retraining, indicating the learned proprioceptive mapping is not overfit to one robot unit.","At roughly 2,000 Hz with an EMA filter, predictions are fast enough for live human-robot interaction such as touch-triggered quadruped choreography and pick-and-place instructions through virtual buttons.","Data collection is short (2.5 hours on Spot, 12 minutes on Franka), so equipping a new robot with touch sensing costs labeling time rather than hardware integration."],"supporting_citations":[{"why":"Introduces the contact particle filter baseline that UniTac positions against as a model-driven alternative requiring expert tuning.","marker":"Manuelli and Tedrake (2016)"},{"why":"Shows momentum-based torque-only contact sensing on an arm; UniTac's virtual-button demonstration is inspired by this line of work.","marker":"Iskandar et al. (2024)"},{"why":"Establishes the data-driven proprioceptive contact localization paradigm on a Jaco arm, the baseline UniTac improves on.","marker":"Zwiener et al. (2018)"},{"why":"Trains a network to localize contact on Franka in simulation; UniTac contrasts its real-world data collection with this sim-to-real approach.","marker":"Liang and Kroemer (2021)"},{"why":"Frames contact localization as regression or classification, the design choice UniTac evaluates.","marker":"Molchanov et al. (2016)"},{"why":"Documents the integration costs of tactile sensors, motivating the sensor-free approach.","marker":"Dahiya et al. (2010)"},{"why":"Surveys tactile human-robot interaction and motivates the application scenarios.","marker":"Argall and Billard (2010)"}],"fun_headline_variants":["Whole-body touch localization with no extra sensors","Joint sensors alone map contact points on robots","Touch sensing from proprioception at 2,000 Hz","No tactile skin? Joint data still finds touches","UniTac: whole-robot touch from joint positions and torques"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that touches applied by a human at a small set of hand-picked body points, with the robot stationary and joint poses randomized in a limited range, are representative of every contact the robot will experience; the held-out test points come from the same discrete set and collection protocol, so the reported errors do not by themselves show accurate localization of arbitrary continuous or moving touches.","fun_headline_variants_meta":{"raw":{"variants":["Whole-body touch localization with no extra sensors","Joint sensors alone map contact points on robots","Touch sensing from proprioception at 2,000 Hz","No tactile skin? Joint data still finds touches","UniTac: whole-robot touch from joint positions and torques"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000243,"raw_usage":{"total_tokens":1506,"prompt_tokens":900,"completion_tokens":606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":529}},"tokens_in":516,"tokens_out":606,"duration_ms":6483,"temperature":1.0,"reasoning_tokens":529,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:27:39.323023+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a Spot or Franka, train UniTac on the paper's sampled contact points, then touch a location between those points (or on a Spot leg, which the paper excluded) while the robot is stationary, with ground truth measured by a motion-capture marker or a reference tactile patch. If the average localization error on these continuous, out-of-distribution touches is substantially larger than the reported 7-8 cm, then the 'whole-body' claim holds only for the discrete, hand-picked contacts the model was trained on.","supporting_citations":[{"cited_title":"Localizing external contact using proprioceptive sensors: The Contact Particle Filter","cited_arxiv_id":null,"evidence_quote":"Introduces the contact particle filter baseline that UniTac positions against as a model-driven alternative requiring expert tuning."},{"cited_title":"Intrinsic sense of touch for intuitive physical human-robot interaction","cited_arxiv_id":null,"evidence_quote":"Shows momentum-based torque-only contact sensing on an arm; UniTac's virtual-button demonstration is inspired by this line of work."},{"cited_title":"Contact Point Localization for Articulated Manipulators with Proprioceptive Sensors and Machine Learning","cited_arxiv_id":null,"evidence_quote":"Establishes the data-driven proprioceptive contact localization paradigm on a Jaco arm, the baseline UniTac improves on."},{"cited_title":"Contact Localization for Robot Arms in Motion without Torque Sensing","cited_arxiv_id":null,"evidence_quote":"Trains a network to localize contact on Franka in simulation; UniTac contrasts its real-world data collection with this sim-to-real approach."},{"cited_title":"Sukhatme","cited_arxiv_id":null,"evidence_quote":"Frames contact localization as regression or classification, the design choice UniTac evaluates."}],"review_version":1}