{"id":"447b8e10-187f-4181-8ab9-d5af6fe5d9e4","arxiv_id":"2607.07287","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A multi-timescale tactile hierarchy with subtask planning, tactile world-model goals, and residual refinement raises real-robot success by about 16–19 points over strong baselines on six contact-rich tasks.","lead":"TouchWorld is a hierarchical robot policy that uses touch both to predict future contact goals and to correct actions at high frequency during dexterous tasks. It reports large success-rate gains over strong vision-language and tactile baselines on six real-robot contact-rich tasks, including under human perturbations.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The headline gains may be driven more by residual refinement than by the full predictive hierarchy the paper claims as the foundation model.","rationale":"The reader's weakest assumption (platform/scheduling transfer) is real and correctly noted in Limitations and Implementation Details, but it is secondary to whether the paper's own evidence attributes the Table 1 gains to the full predictive-reactive hierarchy versus residual refinement alone. Without numeric ablations, the strongest claim is only partially supported even on the evaluated Wuji+JQ stack. That does not overturn CONDITIONAL—real-robot gains and component analyses still make an accept-shaped systems contribution if attribution and reproducibility are tightened—but it is the more load-bearing internal concern. Agreement with the reader is partial: same overall verdict and platform caveats, different primary soft spot (causal attribution of gains vs. generalization of fixed hyperparameters).","tokens_in":16924,"tokens_out":529,"duration_ms":76305,"concrete_test":"From the same 100-rollout protocol as Table 1, report exact average success (clean and perturbation) for: (i) full TouchWorld, (ii) no Tactile World Model, (iii) no Subtask Planner, (iv) no residual refinement, and (v) residual-only on a frozen FTP-1/Pi-0.5 nominal policy. If (iv) or (v) closes most of the 15.7/18.5 pp gap, the hierarchy-as-foundation claim weakens; if the full stack remains clearly best, the claim holds.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim attributes the 15.7/18.5 pp gains over FTP-1 and other baselines to the full multi-timescale stack (Subtask Planner + Tactile World Model + goal-conditioned VLA + high-frequency residual refinement). Table 1 and the abstract present this as a predictive-and-reactive foundation model. However, §4.4 and Fig. 5 only give qualitative stacked ablations: removing residual refinement is said to hurt the perturbation setting most, while removing the world model or planner mainly hurts long-horizon consistency. No numeric per-component success rates, confidence intervals, or matched-compute controls are reported. Because residual correction alone is a known strong lever in contact-rich control (and FTP-1 is already a tactile policy), it remains possible that most of the reported lift is residual feedback on this platform rather than the claimed predictive hierarchy. The foundation-model framing then rests on an incompletely quantified attribution.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"TouchWorld proposes a hierarchical tactile foundation policy for dexterous manipulation that separates multi-timescale roles: a High-Level Planning Layer (Subtask Planner + Tactile World Model) for executable subtasks and predicted visual-tactile subgoals, a Visuo-Tactile Goal-Conditioned Policy that generates nominal action chunks via flow matching, and a Tactile Residual Transformer that applies high-frequency residual corrections from recent tactile and proprioceptive histories. The system is trained in four stages (planner SFT, human-then-robot world-model adaptation, nominal VLA imitation, residual refinement) and evaluated on six real-robot contact-rich tasks in clean and human-perturbation settings. Table 1 reports 65.0% average clean success and 53.7% under perturbation, outperforming Pi-0.5, FTP-1, and GR00T N1.7 by up to 15.7 and 18.5 percentage points; supporting analyses include stacked ablations (Fig. 5), world-model contact metrics (Table 2), and planner correctness/execution metrics (Table 3).","tokens_in":17243,"tokens_out":1046,"duration_ms":132273,"significance":"If the gains hold under stronger attribution and statistical reporting, the paper would be a useful systems contribution: it shows how to keep VLA-style semantic generalization while giving touch both a predictive subgoal role and a fast residual role, rather than treating tactile as another low-rate token stream. Strengths include a real six-task clean/perturbation benchmark, external baselines (including the prior tactile policy FTP-1), staged training that reuses human EgoTouch data, and separate quantitative checks of the world model and subtask planner. The residual-learning formulation is standard rather than circular, and the hierarchy is a concrete, implementable design for contact-rich dexterous control. The main open question is how much of the lift is the full predictive stack versus residual feedback on this platform.","major_comments":[{"comment":"§4.3–4.4, Table 1, Fig. 5: The central claim attributes the 15.7/18.5 pp gains over the strongest baseline (FTP-1) to the full predictive-and-reactive hierarchy. Fig. 5 only provides qualitative stacked ablations without numeric per-component success rates, confidence intervals, or matched-compute controls. Because residual correction is already a strong lever in contact-rich control and FTP-1 is a tactile policy, the manuscript needs quantitative ablation numbers (at least average clean vs. perturbation success for each removed component) so readers can test whether residual refinement alone explains most of the lift.","section":null},{"comment":"§4.1 and Table 1: Each task reports success over 100 rollouts, but no standard errors, binomial CIs, or significance tests are given. With averages of 65.0% and 53.7%, uncertainty on the order of a few percentage points is material for the claimed margins over FTP-1 (49.3%/35.2%). Adding CIs or bootstrap intervals is load-bearing for the headline comparison.","section":null},{"comment":"§5 Limitations and Appendix B: The residual subspace is restricted to 58 of 120 action dimensions, with fixed H=32, W=16, C=4 and an image-form tactile interface tied to the Wuji hand + JQ glove. The foundation-model framing in the abstract and contributions should be tempered, or supported by at least one transfer/sensitivity experiment, so the claim is not overstated relative to a platform-specific multi-rate stack.","section":null}],"minor_comments":[{"comment":"§2, Eqs. (1)–(4): Define H, W, k, C, and the residual subspace earlier in the main text; several appear only later in §4.1 or Appendix B.","section":null},{"comment":"Fig. 5: The stacked bars are hard to read for per-task contribution; a companion numeric table would help.","section":null},{"comment":"Table 2: State the pressure threshold τ used for contact/volumetric IoU.","section":null},{"comment":"§3.2 / Appendix B: Clarify how EgoTouch pressure maps are converted into the same visual-tactile goal grids as the robot glove (normalization, layout remapping).","section":null},{"comment":"Related Work: A short explicit comparison to concurrent reactive tactile / residual VLA lines (e.g., Reactive Diffusion Policy, T-Rex, FTP-1) would sharpen the novelty claim.","section":null},{"comment":"Date line says July 10, 2026 while arXiv stamp is 9 Jul 2026; align for consistency.","section":null}],"recommendation":"major_revision","confidential_remarks":"The empirical story is real-robot and stronger than many VLA+tactile systems papers, but the foundation-model language and incomplete ablation quantification are the main risks for a top venue. If the authors supply numeric ablations and CIs, this could become a solid accept after revision; without them the residual-only alternative remains too plausible."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: they built a working multi-rate stack that keeps a VLA for semantics, a short-horizon tactile world model for contact goals, and a high-frequency residual layer for slip/insertion recovery, and it beats Pi-0.5, FTP-1, and GR00T N1.7 by a clear margin on six contact-rich tasks (65% clean, 53.7% under human perturbation).\n\nWhat is actually new is the integration and the measurement, not any single module. Subtask planning, action chunking, tactile world models, and residual tactile control all exist; putting them on separate timescales with staged training and reporting clean vs. perturbation numbers is the contribution. Table 1 is consistent across tasks. Tables 2–3 and the planner analysis are honest extras: the world model beats copy/nearest-neighbor on contact IoU, and memory-augmented SFT beats a larger zero-shot planner. Residual targets as demo-minus-nominal is standard, not circular. Citations look fair; they name FTP-1, T-Rex, reactive diffusion, and hierarchical VLA work rather than pretending isolation.\n\nSoft spots, in proportion: the stress-test has a point. Ablations in Fig. 5 are stacked and qualitative—no numeric per-component rates, no CIs on 100 rollouts, no matched-compute residual-only control—so you cannot cleanly attribute the 15–18 pp to the full predictive hierarchy versus residual feedback on this glove/hand. Limitations already admit platform-specific sensing, fixed H/W/C schedules, and a six-task suite. No public code/data. That makes the “foundation model” label aspirational; the systems result still stands.\n\nThis is for people building contact-rich VLAs or multi-rate tactile controllers. Worth a serious referee. I would engage: cite the hierarchy and the clean/perturbation protocol, and treat the residual-vs-world-model split as open until someone re-runs the ablations with numbers.","headline":"Solid multi-rate tactile systems paper with real-robot gains; the hierarchy is useful, but residual refinement may carry more of the lift than the foundation-model framing admits.","tokens_in":17903,"tokens_out":520,"would_cite":true,"duration_ms":6776,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"TouchWorld shows that treating touch as both a predicted contact goal and a fast residual correction signal, on separate timescales from vision-language planning, raises success on long-horizon contact-rich robot tasks.","keywords":["tactile foundation model","dexterous manipulation","hierarchical policy","tactile world model","residual refinement","vision-language-action","contact-rich manipulation","real-robot benchmark"],"falsifier":"On the same six tasks with matched data and sensing, a strong monolithic tactile policy that fuses touch at the same rate as vision-language tokens would match or exceed TouchWorld's clean and perturbation success rates, or ablations removing residual refinement and tactile world-model goals would not drop performance as reported.","tokens_in":17797,"feed_emoji":"🤖","tokens_out":895,"duration_ms":14697,"temperature":0.7,"pith_summary":"Everyday dexterous skills fail when a robot cannot anticipate how contact should evolve and cannot quickly fix slip, misalignment, or force errors. Vision and language give semantic guidance but hide those contact states; most policies still fold touch into one slow monolithic action loop. TouchWorld claims that a multi-rate hierarchy fixes this: a slow planner emits executable subtasks and tactile subgoals, a mid-rate visuo-tactile policy generates nominal action chunks, and a high-rate residual policy corrects those chunks from recent tactile and proprioceptive feedback. On six real-robot tasks spanning watering, clearing, insertion, wiping, and soft-object pulling, this design reaches 65.0% average success clean and 53.7% under human perturbation, beating the strongest baseline by 15.7 and 18.5 points. A sympathetic reader cares because the result says foundation robot policies can keep language-level generalization while becoming robust at contact if touch is given both a predictive and a reactive role.","feed_headline":"Hierarchical touch lifts contact-task success by 16 points","feed_subtitle":"Predict contact goals slowly, correct slip and force errors fast, and long-horizon robot skills hold under perturbation.","key_machinery":"TouchWorld hierarchy: High-Level Planning Layer (Subtask Planner plus Tactile World Model) supplies executable subtasks and predicted visual-tactile subgoals; Visuo-Tactile Goal-Conditioned Policy emits nominal action chunks; Tactile-Conditioned Refinement Policy (Tactile Residual Transformer) adds online residuals from high-frequency tactile and proprioceptive history.","core_discovery":"The paper establishes that a predictive-and-reactive tactile foundation model, implemented as a hierarchical multi-timescale policy, improves long-horizon contact-rich dexterous manipulation by separating vision-language subtask planning, tactile world-model subgoal prediction, visuo-tactile goal-conditioned action generation, and high-frequency tactile residual refinement, rather than coupling them in a single monolithic loop.","pith_inferences":["If residual commit interval and world-model refresh rules were made contact-state adaptive, compute cost could fall during stable contact phases without losing recovery speed.","The same predictive-plus-reactive split may help force or torque streams on rigid grippers even without high-resolution tactile images.","Uncertainty-aware multi-hypothesis tactile subgoals could address the paper's own limit on longer-horizon contact futures under occlusion or human disturbance.","Sensor-layout transfer may reduce to shared image-form tactile goals plus a small residual adapter rather than full policy retraining."],"forward_implications":["Contact-rich foundation policies should keep semantic planning slow while dedicating a separate high-rate path to tactile residual correction.","Predicted tactile subgoals can condition nominal action chunks without forcing the full model to replan at contact rates.","Human bimanual tactile pretraining plus robot fine-tuning can supply contact priors that improve robot tactile subgoal prediction.","Perturbation robustness becomes a first-class evaluation axis for tactile VLA systems, not an afterthought.","Modular fallbacks (task prompt only, no predicted goals) remain usable when the planner or world model is unavailable."],"fun_headline_variants":["TouchWorld hierarchy boosts dexterous contact tasks by 16 points","Predictive tactile model separates planning from fast contact correction","Hierarchical touch policy raises long-horizon robot success 16 pts","Tactile world model predicts goals and corrects slip in real time","Multi-timescale touch lifts perturbed manipulation success by 18 pts"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The claim assumes that the fixed multi-rate schedule, image-form tactile interface, and residual action subspace used on this hand-and-glove setup will transfer as a general foundation rather than remaining a platform- and task-specific stack.","fun_headline_variants_meta":{"raw":{"variants":["TouchWorld hierarchy boosts dexterous contact tasks by 16 points","Predictive tactile model separates planning from fast contact correction","Hierarchical touch policy raises long-horizon robot success 16 pts","Tactile world model predicts goals and corrects slip in real time","Multi-timescale touch lifts perturbed manipulation success by 18 pts"]},"model":"grok-4.5","effort":"low","cost_usd":0.003732,"raw_usage":{"total_tokens":1254,"prompt_tokens":852,"num_sources_used":0,"completion_tokens":90,"cost_in_usd_ticks":37320000,"prompt_tokens_details":{"text_tokens":852,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":312,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":852,"tokens_out":90,"duration_ms":3380,"temperature":1.0,"reasoning_tokens":312,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T19:38:30.510079+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the same six tasks with matched data and sensing, a strong monolithic tactile policy that fuses touch at the same rate as vision-language tokens would match or exceed TouchWorld's clean and perturbation success rates, or ablations removing residual refinement and tactile world-model goals would not drop performance as reported.","supporting_citations":[],"review_version":2}