{"id":"8a6ba079-6187-43d8-b6f0-441ed1737ec7","arxiv_id":"2510.20390","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A robot grasping system that uses neural descriptor fields to guide a tactile RL policy achieves sub-millimetre precision in simulation and zero-shot transfer to real peg-in-hole and bottle-lid-opening tasks.","lead":"This paper combines vision-based neural descriptors with tactile sensing and reinforcement learning to make robot grasping more precise, then shows the system works on real plugs, USB sticks, and bottle lids without extra training. If it holds up, it gives a recipe for teaching robots touch-guided manipulation from a few demonstrations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Zero-shot claim is only demonstrated on cylindrical features; the paper's own tactile-transfer failure on curved surfaces and light contacts undermines the 'general framework' generalization.","rationale":"We identify the same load-bearing assumption as the reader: the real-to-sim tactile transfer must be accurate enough for zero-shot policy deployment across the range of geometries and contact forces the method claims to handle. The paper's own text provides the critical evidence against that assumption in its current form: IV-C3 explicitly concedes poor transfer on curved mug/handle surfaces and V-C1 concedes light-contact failures. Because the real-world evaluation set is deliberately limited to cylindrical objects, the headline claim of a 'general framework' is not yet supported in the positive direction — the negative result for curved surfaces is a direct counterexample to the unrestricted version of the claim. This is a genuine limitation but not a fatal flaw: the method is clearly effective on the tested cylindrical tasks, and the simulation ablation supports the NDF-conditioned policy design. We therefore keep the reader's CONDITIONAL verdict. The proposed test — a real-world non-cylindrical feature run — would settle whether the limitation is a boundary condition or a fundamental barrier.","tokens_in":14346,"tokens_out":3736,"duration_ms":37481,"concrete_test":"Run the full NeuralTouch pipeline zero-shot on a non-cylindrical target such as a mug rim or horizontal handle, using a pix2pix transfer model trained on curved-surface data (or the same transfer model if possible). If real-world success for these features is comparable to the simulation success rates (≥90% in Table II), the generality concern is resolved; if success falls significantly below, the zero-shot claim is limited to cylindrical contacts. As a secondary check, record contact forces and report success on light-contact trials to test the aliasing failure mentioned in Sec. V-C1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim's 'general framework' depends on the real-to-sim tactile image transfer (pix2pix GAN) being accurate across object geometries and contact forces. That condition is the least secure part of the paper. Section IV-C3 states the transfer 'performed unsatisfactorily with the curved surfaces of the mugs and handles,' so real-world experiments were restricted to cylindrical features. Section V-C1 further notes that light contacts produce incorrect simulated images because marker motion is too subtle, and attributes this to tactile aliasing. Consequently, every real-world zero-shot result (bottle lids; bolt/plug/USB peg-out/in-hole) exercises only cylinder-like geometry with relatively firm contacts. The simulated successes on mug rims and handles are not zero-shot transferred, so they do not count as evidence for the real-world generality claim. This is a boundary condition, not an internal contradiction: the method may work more broadly, but the only explicit test of non-cylindrical geometry in reality was negative. The abstract's 'general framework for precise, contact-rich robotic manipulation' therefore goes beyond the reported evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NeuralTouch, a two-stage manipulation framework that combines Neural Descriptor Fields (NDF) for coarse, category-level grasp pose generation with a deep RL policy that refines the pose using tactile images, proprioception, and NDF target descriptors. The fine-phase policy is trained in simulation with PPO and transferred zero-shot to a real Franka arm with TacTip sensors via a pix2pix-based real-to-sim tactile translation. In simulation, NeuralTouch reduces position errors from 9.5-13.6 mm (NDF) to 0.7-1.0 mm and outperforms an RL-touch baseline lacking descriptors. Real-world experiments on bottle-lid opening and peg-out/in-hole with cylindrical objects report higher success rates than NDF alone. The paper claims this is a general framework for precise contact-rich manipulation.","tokens_in":14597,"tokens_out":6046,"duration_ms":54793,"significance":"If the reported results hold, NeuralTouch is a valuable contribution: it shows that conditioning a tactile RL policy on neural descriptors can substantially improve precision over vision-only NDF while retaining category-level generalization, and it demonstrates zero-shot sim-to-real for cylindrical features. The simulation ablation is well motivated, the online target-switching result (Fig. 6) is compelling, and the authors are transparent about the tactile-transfer limitations on curved surfaces and light contacts. However, the absence of statistical rigor and the narrow real-world geometry scope mean the evidence does not yet support the strong 'general framework' claim.","major_comments":[{"comment":"The central quantitative claims are reported without error bars, standard deviations, or trial counts. Table I gives 'average' errors with no variance despite the training curves in Fig. 5 being over 10 seeds; Table II gives success rates for 60 trials without confidence intervals; Tables III-IV give success rates without stating the number of real-world trials. This makes it impossible to assess whether the large improvements (e.g., 0.8 mm vs 13.6 mm in Table I) are significant or stable. Please report per-seed/per-trial statistics, confidence intervals, and trial counts for all tables.","section":"Sec. V-A / Tables I-IV"},{"comment":"The training initial poses are sampled uniformly from [±20 mm, ±20°] rather than from the NDF optimizer, because 'initial grasping poses are not sampled from the NDF optimizer during training' (Sec. IV-C2). The Table I comparison with the NDF baseline is therefore not an end-to-end comparison from the same coarse-pose distribution: the NDF row reports errors of the NDF-generated pose, while the NeuralTouch row reports errors after refinement from random poses within a fixed box. To substantiate the claim that NeuralTouch improves NDF-based grasping, the evaluation should initialize the policy from the NDF output (or a held-out set of NDF failures) and report the resulting final errors, or at least report the distribution of NDF initial errors and show the policy is robust over that distribution.","section":"Sec. IV-C2 / V-A / Eq. (4)"},{"comment":"The zero-shot sim-to-real evidence is restricted to cylindrical features. The authors state that the tactile real-to-sim transfer 'performed unsatisfactorily with the curved surfaces of the mugs and handles' (Sec. IV-C3), so real-world experiments were limited to cylinders; they additionally report failures on light contacts due to tactile aliasing (Sec. V-C1). Consequently, the real-world results support the method for cylinder-like geometries with relatively firm contacts only. The abstract's 'general framework for precise, contact-rich robotic manipulation' and the phrase 'various objects' overstate the demonstrated scope. Either add real-world experiments on non-cylindrical features or revise the claims to specify cylindrical-contact geometry.","section":"Sec. IV-C3 / V-C1 / Abstract"},{"comment":"The real-to-sim tactile image transfer is a load-bearing component of the zero-shot claim, but it is evaluated only qualitatively. No quantitative validation of the pix2pix transfer accuracy (e.g., pose-estimation error on a held-out set of real contacts) is provided, and the paper notes that it 'performed unsatisfactorily' for curved surfaces. Given that the policy depends on this transfer, please report quantitative transfer accuracy and, ideally, analyze how transfer error compares to the policy's tolerance. This would clarify the boundary conditions of the otherwise promising results.","section":"Sec. IV-C3 / Fig. 2"}],"minor_comments":[{"comment":"Typo: 'relared' should be 'related'.","section":"Sec. II-C"},{"comment":"The number of target descriptor samples per feature is fixed at 12, but the sampling procedure is not described; please specify how the 12 descriptors are drawn (e.g., from demonstrations, perturbed poses) and whether this choice affects performance.","section":"Sec. IV-C1 / Eq. (4)"},{"comment":"The caption labels 'Without Neural Descriptors' and 'With Neural Descriptors' without indicating which curve corresponds to which color; please clarify in the caption or legend.","section":"Fig. 5"},{"comment":"The heading 'average error' is ambiguous because the rows aggregate over six target features; please define how the average is computed and report per-feature results with variance.","section":"Table I"},{"comment":"The online target-switching demo is interesting, but the quantitative success of switching (e.g., number of trials) is not reported; please add at least a brief evaluation or clarify that it is a qualitative illustration.","section":"Sec. III-C / Fig. 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound in its core idea, but the presentation currently overclaims generality. The main risks are (1) missing statistical detail, (2) the train/test initial-pose mismatch relative to the NDF optimizer, and (3) the restriction of real-world validation to cylindrical features. If the authors can address these, the paper could be a solid contribution. I would not recommend rejection, as the limitations are fixable through additional analysis and a more measured framing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my take on NeuralTouch. The core contribution is real. Conditioning a single tactile RL policy on NDF descriptors, so the policy can adapt to different target contact geometries without retraining, is a clean step beyond earlier tactile servoing work that locks onto one predefined contact shape. The simulation ablations are convincing—position errors drop from ~10–20 mm to under 1 mm, and the online descriptor-switching demo is a nice bonus. The real-world bottle-lid and peg-out/in-hole results beat the NDF baseline by a wide margin, even if absolute success rates on the tighter clearances are modest. The paper is honest about its limitations, which I respect.\n\nThe soft spots are the ones you'd expect. The real-to-sim tactile transfer (pix2pix GAN) is the weak link, and the paper admits it in Section IV-C3: the transfer 'performed unsatisfactorily with the curved surfaces of the mugs and handles.' So every real-world zero-shot test in this paper is on cylindrical features. The simulated successes on mug rims and handles don't transfer, so the 'general framework' claim in the abstract isn't backed by the real-world evidence. That's a boundary condition, not a fatal flaw, but it's the main thing a referee should push on. Also, Tables I, III, and IV have no error bars or trial counts—just single percentages—so we can't judge variance. The training initial-pose distribution is a fixed box, not the NDF optimizer, so the sim-to-real mismatch is somewhat under-tested; that's minor given the real transfer worked.\n\nThe stress-test note is correct: the zero-shot claim needs to be restricted to what was actually shown, or the authors need to add real-world tests on non-cylindrical geometry. This is fixable in revision.\n\nBottom line: this deserves peer review. It's a solid, empirical result with honest discussion. I'd cite it for the idea of descriptor-conditioned tactile RL, and I'd want to see a revised version that either narrows the generality claim or broadens the real-world experiments.","headline":"NeuralTouch shows that NDF-conditioned tactile RL is a genuine improvement for precise grasping, but the real-world zero-shot evidence only covers cylindrical features—the abstract's 'general framework' language outruns the data.","tokens_in":15076,"tokens_out":2363,"would_cite":true,"duration_ms":21702,"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":"NeuralTouch claims that a tactile reinforcement learning policy conditioned on neural descriptor fields can refine vision-based grasping poses to sub-millimeter accuracy and transfer zero-shot to real robots.","keywords":["neural descriptor fields","tactile servoing","sim-to-real transfer","reinforcement learning","robot grasping","peg-in-hole","multimodal manipulation","tactile sensing"],"falsifier":"Take a previously unseen object with a curved or non-cylindrical contact surface (e.g., a mug handle or a spherical knob), run the NeuralTouch pipeline with the same trained policy and translation model, and measure the achieved grasp pose error and task success. If the success rate collapses or errors exceed a few millimeters, the claim that the framework generalizes beyond cylindrical features is falsified.","tokens_in":14242,"feed_emoji":"🖐️","tokens_out":3714,"duration_ms":30793,"temperature":0.7,"pith_summary":"NeuralTouch combines visual Neural Descriptor Fields (NDF) with tactile reinforcement learning to refine robot grasping poses. The paper claims that conditioning a tactile RL policy on NDF pose descriptors lets a single policy grasp different contact features—rims, handles, bottle necks, bolt heads—without predefined contact geometry. In simulation, grasp position errors drop from 11–14 mm with NDF alone to under 1 mm. The same policy transfers zero-shot to real robots, improving bottle-lid opening from 30–45% to 85–90% success and enabling peg-out/in-hole insertions that the vision-only baseline almost never completes. The aim is a general framework for precise, contact-rich manipulation that avoids task-specific contact models.","feed_headline":"NeuralTouch cuts grasp error to under 1 mm","feed_subtitle":"Vision gives a rough pose; touch refines it, and the policy transfers zero-shot to real robots.","key_machinery":"The key object is the neural pose descriptor field: a function that maps a 6D pose and an object point cloud to a descriptor vector, built from an SE(3)-equivariant occupancy network. It defines a target contact pose implicitly, so the RL policy can read the desired geometric relationship without a categorical contact type. The policy is a deep RL network trained with a policy-gradient algorithm, conditioned on the descriptor along with tactile images and proprioception. The descriptor does the work of disambiguating which feature is the target—something tactile feedback alone cannot do, as the ablation with a tactile-only RL baseline shows.","core_discovery":"The central claim is that neural descriptor fields can serve as the conditioning signal for a tactile servoing policy, replacing the hand-specified contact geometry used by prior tactile RL. The policy takes tactile images, proprioception, and an NDF descriptor of the target pose, and outputs a 7D action (end-effector twist and gripper width). Because NDF descriptors are SE(3)-equivariant and encode geometric correspondence across objects in a category, the same trained policy can servo to different target features on unseen objects, and even switch targets online by updating the input descriptor. The authors show that this raises simulated grasping accuracy to roughly a millimeter position","pith_inferences":["If the descriptor continues to generalize, combining NeuralTouch with closed-loop insertion policies—which the authors note are complementary—could yield full assembly tasks that are robust to pose errors.","The tactile real-to-sim transfer is the bottleneck; improving its fidelity for light contacts and curved surfaces, perhaps with physics-based simulators or better domain randomization, would likely extend the method beyond cylindrical features.","Because the descriptor can be swapped online, the same policy architecture could be trained once on many categories and then used for few-shot adaptation to new tasks by providing one or a few demonstrations of the target pose descriptor.","The failure mode of tactile aliasing on light contacts suggests that adding force or shear information, or temporal accumulation of contacts, may be necessary for sub-millimeter real-world accuracy."],"forward_implications":["Grasping accuracy in simulation improves to sub-millimeter position error (0.7–1.0 mm across six target features), compared to 9.5–13.6 mm for NDF alone.","A single policy, without retraining, can servo to different target contact features and switch targets online by changing the input descriptor.","Zero-shot sim-to-real transfer works for cylindrical features: bottle-lid opening success rises to 85–90% and peg-out/in-hole reaches 55% for a bolt with 2 mm clearance.","The framework removes the need for predefined contact-geometry types in tactile servoing, extending from flat edges and surfaces to curved surfaces and handles."],"fun_headline_variants":["NDF-guided tactile policy hits millimeter grasp accuracy","Tactile RL refined by neural descriptors cuts grasp error","NeuralTouch: zero-shot tactile servoing to sub-mm poses","Descriptor-conditioned touch policy transfers to real robots"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire zero-shot sim-to-real claim rests on the real-to-sim tactile image translation being accurate enough—and the paper itself reports that this transfer fails on curved surfaces and on very light contacts, so if that translation is inaccurate for a new geometry or contact force, the policy cannot transfer.","fun_headline_variants_meta":{"raw":{"variants":["NDF-guided tactile policy hits millimeter grasp accuracy","Tactile RL refined by neural descriptors cuts grasp error","NeuralTouch: zero-shot tactile servoing to sub-mm poses","Descriptor-conditioned touch policy transfers to real robots"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":9.9e-05,"raw_usage":{"total_tokens":845,"prompt_tokens":730,"completion_tokens":115,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":50}},"tokens_in":474,"tokens_out":115,"duration_ms":1756,"temperature":1.0,"reasoning_tokens":50,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T08:26:09.479497+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a previously unseen object with a curved or non-cylindrical contact surface (e.g., a mug handle or a spherical knob), run the NeuralTouch pipeline with the same trained policy and translation model, and measure the achieved grasp pose error and task success. If the success rate collapses or errors exceed a few millimeters, the claim that the framework generalizes beyond cylindrical features is falsified.","supporting_citations":[],"review_version":1}