{"id":"5340de07-f4b3-4b09-9070-739f754ed72a","arxiv_id":"2607.03723","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A policy-agnostic two-stage real-world RL method learns tactile residual corrections on frozen visual policies, lifting contact-rich task success from 5–40% to 85–100% in under 80 minutes.","lead":"OmniTacTune adapts tactile sensing to frozen visual robot policies via a two-stage real-world residual RL pipeline, without offline tactile demos. It raises success on four contact-rich tasks from 5–40% to 85–100% in 40–80 minutes of practice.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Success gains may be overstated by sparse evaluation N and human terminal labels without variance or automatic success detection.","rationale":"The reader correctly flags the bootstrap (base rollouts + ControlTac + multi-sensory reward + human terminal labels) as the weakest assumption and already assigns CONDITIONAL pending statistical reporting. My concern is the same soft spot made concrete: the absolute success numbers that carry the paper are measured with N=10–20 and human success labels, so the claimed 85–100% and the large baseline gaps are not yet statistically or operationally locked down. The systems idea (policy-agnostic tactile residual RL) remains well-supported by the ablations and cross-policy/tactile results; the issue is only whether the headline efficiency numbers are robust enough for an unconditional accept. Hence verdict stays CONDITIONAL, agreement with the reader is full, and the concrete test is a higher-N automatic-success re-evaluation that would settle the point without requiring new training runs.","tokens_in":27925,"tokens_out":564,"duration_ms":5497,"concrete_test":"Re-evaluate the final OmniTacTune and PLD* policies on each of the four tasks with N≥50 independent trials (fixed initial-pose randomization), report success rate ± 95% Wilson CI or bootstrap SE, and replace human terminal labels with an automatic geometric success detector (e.g., peg fully seated / charger fully inserted / cap lifted past threshold / box lid angle). If any task’s lower CI falls below ~70% or the gap vs PLD* shrinks below 20 points, the strongest claim needs qualification.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (5–40% → 85–100% in 40–80 min; Table 1, Fig. 4) rests on final success rates from only 20 trials per task and intermediate checkpoints from 10 trials, with no reported variance, confidence intervals, or multi-seed runs. Terminal reward is a human-assigned success label of 1 (Sec. 3.4, App. A.6), so the same operator who resets the robot also defines the learning signal and the evaluation metric. For Cap Opening and Box Opening the base starts at 5%; if warm-start rarely produces near-contact states, residual learning depends heavily on the hand-designed flow + tactile reward and on those sparse human labels. The large margins over PLD*/ViTAL could therefore partly reflect evaluation noise or label consistency rather than a reliably superior residual pipeline. This is the softest load-bearing point under the reader’s weakest assumption: without denser, automatic success detection and statistical reporting, the headline efficiency claim is not fully secured.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"OmniTacTune proposes a policy-agnostic two-stage real-world RL pipeline that adapts tactile feedback to frozen visual base policies via residual correction, without offline tactile demonstrations. Stage 1 warm-starts a flow-tactile critic and tactile encoder from autonomous base-policy rollouts (with ControlTac trajectory-level tactile augmentation); Stage 2 learns a lightweight residual actor that adds contact-gated corrections on top of the base action chunk, guided by a multi-sensory reward combining object-centric flow subgoals, tactile grasp/safety terms, and human terminal success labels. On four contact-rich tasks (peg-in-hole, charger insertion, cap opening, box opening), the method reports improving base success from 5–40% to 85–100% in 40–80 minutes, with generalization across base policies (human/teleop flow, ACT, DP, π0.5) and tactile representations (AnyTouch2, Sparsh, T3, markers), outperforming adapted PLD and ViTAL baselines and visuo-tactile imitation under a matched time budget.","tokens_in":28279,"tokens_out":905,"duration_ms":14224,"significance":"If the reported efficiency and generality hold under stronger evaluation, this is a practically important contribution: it offers a concrete path to attach tactile residual practice to scalable visual priors rather than collecting large paired visuo-tactile datasets. Strengths include real hardware results on four genuinely contact-rich tasks, systematic compatibility tests across policy architectures and tactile encoders, relevant residual-RL and imitation baselines, and ablations of reward components, residual design, warm-start, and action scaling (Sec. 4, App. C). The residual interface (shared flow goals + base action chunks + contact gate) is a clean systems idea for policy-agnostic adaptation. These are falsifiable empirical claims with public project-page materials, not circular constructions.","major_comments":[{"comment":"Table 1 and Fig. 4 (also Sec. 4.1): the central efficiency claim (5–40% → 85–100% in 40–80 min) rests on final success from 20 trials and intermediate checkpoints from 10 trials, with no variance, confidence intervals, multi-seed runs, or multi-operator evaluation. For Cap Opening and Box Opening the base starts at 5%, so absolute margins over PLD*/ViTAL are large but statistically under-specified. Please report binomial CIs or bootstrap intervals at minimum, and preferably repeated training runs or denser evaluation; without this the headline numbers are not fully secured.","section":null},{"comment":"Sec. 3.4 and App. A.6: terminal success is a human-assigned reward of 1, and the same operator performs resets. This couples the learning signal and the evaluation metric. Clarify whether evaluation success uses the same human judgment as training, whether any automatic success detector exists, and how label consistency was controlled across methods. If evaluation is fully human-labeled, consider a blinded protocol or automatic geometric/contact criteria so reported gains cannot be attributed to label drift.","section":null},{"comment":"Sec. 3.3–3.4 and the weakest operating regime (5% base policies): residual learning assumes warm-start rollouts plus flow/tactile shaping produce enough near-contact experience for a usable critic. The paper shows warm-start ablations (Fig. 17) but does not quantify contact-state coverage or success of warm-start trajectories on Cap/Box Opening. Please report how often base rollouts reach contact/near-goal states, and discuss failure modes when the base rarely enters the residual’s useful region—this is load-bearing for the claim of no offline tactile demos.","section":null},{"comment":"Sec. 4.3 / Table 2: imitation baselines receive 40 extra teleop demos under a 50-minute budget, while OmniTacTune uses online interaction with human terminal labels and dense shaped rewards. The comparison is informative but not fully matched in supervision type. Explicitly state what human effort (resets, success labeling, interventions) is required for OmniTacTune versus teleop collection, so data-efficiency claims are not overstated relative to pure demonstration methods.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: they show you can bolt tactile residual practice onto frozen visual policies (flow, ACT, DP, π0.5) without offline tactile demos, and get large real-world gains on four contact-rich tasks in under 80 minutes.\n\nWhat is actually new is the packaging, not any single module. Residual RL, warm-start rollouts, object-centric flow rewards, and tactile encoders already exist. The contribution is a coherent two-stage pipeline that (1) uses autonomous base rollouts plus ControlTac-style tactile augmentation to bootstrap a flow-tactile critic and encoder, then (2) learns a lightweight residual actor with a contact gate and multi-sensory reward. That is a practical answer to the scale gap between visual priors and scarce tactile data. The experiments are the right ones: four hard tasks (peg, charger, cap, box), multiple base policies, multiple tactile representations (AnyTouch2, Sparsh, T3, markers), and ablations on reward, residual design, warm-start, and action scale. Table 1 and Fig. 4 show clear margins over adapted PLD and ViTAL. The human-flow smoothness analysis in the appendix is honest about why teleop is a weak prior for contact.\n\nSoft spots, in proportion: final success is 20 trials per task, checkpoints 10, no variance or multi-seed runs. Terminal success is a human label of 1 from the same operator who resets. For cap/box the base starts at 5%, so warm-start quality and the hand-designed flow+tactile reward carry a lot of weight. That does not kill the result—the absolute gains and ablations still look real—but it means the headline 5–40% → 85–100% should be read as strong systems evidence, not a statistically locked efficiency law. Reward weights, residual scale schedule, and contact thresholds are free design choices; that is normal for real-world RL, not circularity. Citations are appropriate; self-cites to GenFlowRL/MimicTouch/ControlTac are prior building blocks, not padding.\n\nThis is for people who actually run contact-rich robots and care about attaching touch to scalable visual policies. It deserves a serious referee. I would engage with it, cite the pipeline idea, and push for variance reporting and clearer success automation if it lands in a venue.","headline":"Solid systems paper: residual tactile RL on frozen visual policies works across four real contact tasks and several base policies; the efficiency claim is real but rests on small-N human-labeled success rates.","tokens_in":28851,"tokens_out":599,"would_cite":true,"duration_ms":6183,"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":"OmniTacTune lifts weak visual robot policies to high contact-rich success by learning tactile residual corrections in 40–80 minutes of real-world practice.","keywords":["visuo-tactile policy","real-world RL","contact-rich manipulation","tactile residual adaptation","visual base policies","multi-sensory reward","residual RL","tactile sensing"],"falsifier":"On the same four contact-rich tasks and wall-clock budgets, residual training after the paper’s warm-start fails to raise final success above the visual-only residual baseline and above the frozen base across several base-policy architectures and tactile encodings.","tokens_in":28853,"feed_emoji":"🤖","tokens_out":1092,"duration_ms":17051,"temperature":0.7,"pith_summary":"Vision-only robot policies trained from human video, teleoperation, or demos give scalable motion priors, but they still fail when success hinges on local force and contact geometry that cameras cannot measure. Touch supplies those signals, yet tactile data are scarce and hard to transfer across sensors and tasks. OmniTacTune freezes a pretrained visual policy and adapts tactile feedback as residual correction via real-world RL: a warm-start stage bootstraps a tactile encoder and critic from autonomous base-policy rollouts (with force-conditioned tactile augmentation), then online residual RL learns contact-gated corrections under a multi-sensory reward. On four hardware tasks—peg-in-hole, charger insertion, cap opening, and box opening—it raises success from 5–40% to 85–100% within 40–80 minutes, and the same recipe works across flow, ACT, diffusion, and VLA-style bases and across tactile image and marker representations. A reader who wants scalable vision plus reliable contact cares because this is an efficient path to add touch without collecting large paired visuo-tactile datasets from scratch.","feed_headline":"Tactile residuals lift visual robots to 85–100% in under 80 min","feed_subtitle":"Two-stage real-world RL adapts touch to frozen vision policies without offline tactile demos.","key_machinery":"Two-stage tactile residual RL: Stage 1 warm-starts a flow-tactile critic and tactile encoder from frozen base-policy rollouts (contact-only encoder updates plus trajectory-level force-conditioned tactile augmentation); Stage 2 freezes the base and trains a residual actor a = a_base + s_t a_residual, gated by contact, conditioned on proprioception, object-centric flow goals, tactile features, and the base action chunk, with SAC under a multi-sensory reward of reaching, grasp, flow-subgoal, and safety terms.","core_discovery":"OmniTacTune shows that tactile sensing can be adapted to frozen, architecturally diverse visual base policies through residual real-world RL without offline tactile demonstrations. Autonomous base-policy rollouts warm-start a flow-tactile critic and task-adapted tactile encoder; online residual RL then learns lightweight contact-aware corrections on top of the base actions under object-centric multi-sensory reward shaping. Across four contact-rich real-world tasks this raises success from 5–40% to 85–100% in 40–80 minutes and generalizes across base policies and tactile representations.","pith_inferences":["If the base policy rarely reaches near-contact states, residual learning can stall; the method implicitly needs a prior that already lands in a useful contact neighborhood.","The vision-for-planning, touch-for-refinement split may extend to multi-finger dexterity and bimanual assembly where tactile data remain scarce relative to vision.","Manual resets and human terminal success labels remain practical bottlenecks; automating both would be a direct stress test of the same residual recipe.","Dynamic levering failures (slip, edge miss, pose tilt) suggest residual scale and force-aware rewards may need richer contact profiles than the current scheduler alone provides."],"forward_implications":["Scalable visual policies can be made contact-aware without retraining them or collecting large paired visuo-tactile datasets.","Tens of minutes of real-world residual practice can close the last-mile contact gap that pure imitation leaves open.","One residual interface can attach to flow, ACT, diffusion, and VLA-style bases, so tactile adaptation need not be architecture-specific.","Both compact marker signals and pretrained tactile image encoders can serve as the tactile stream for residual correction.","Warm-start of critic and tactile encoder, plus multi-sensory reward shaping, is material to sample-efficient tactile residual RL."],"fun_headline_variants":["Tactile residuals lift visual policies from 5–40% to 85–100% in 40–80 min","Residual RL adapts touch to frozen vision policies without tactile demos","OmniTacTune residual touch raises contact-rich success to 85–100%","Policy-agnostic real-world RL adds tactile residuals for 85–100% success","Two-stage residual RL boosts visual bases to 85–100% on contact tasks"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Autonomous rollouts of a still-weak visual policy, plus synthetic tactile augmentation and a hand-designed multi-sensory reward with human success labels, are enough to bootstrap a stable critic and residual policy without offline tactile demos.","fun_headline_variants_meta":{"raw":{"variants":["Tactile residuals lift visual policies from 5–40% to 85–100% in 40–80 min","Residual RL adapts touch to frozen vision policies without tactile demos","OmniTacTune residual touch raises contact-rich success to 85–100%","Policy-agnostic real-world RL adds tactile residuals for 85–100% success","Two-stage residual RL boosts visual bases to 85–100% on contact tasks"]},"model":"grok-4.5","effort":"low","cost_usd":0.007854,"raw_usage":{"total_tokens":1913,"prompt_tokens":810,"num_sources_used":0,"completion_tokens":118,"cost_in_usd_ticks":78540000,"prompt_tokens_details":{"text_tokens":810,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":985,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":810,"tokens_out":118,"duration_ms":7030,"temperature":1.0,"reasoning_tokens":985,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T00:22:07.471648+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the same four contact-rich tasks and wall-clock budgets, residual training after the paper’s warm-start fails to raise final success above the visual-only residual baseline and above the frozen base across several base-policy architectures and tactile encodings.","supporting_citations":[],"review_version":1}