{"id":"d765e351-d397-4f58-a695-f81b58a58860","arxiv_id":"2505.04978","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A hierarchical, model-based, contact-implicit controller with tactile tracking completes five real-world in-hand manipulation tasks without any learning, under disturbances.","lead":"This paper presents a two-level robotic system where one module plans finger movements and contact forces in real time, and a second module tracks both using touch feedback, letting a robotic hand rotate, slide, and open objects. A smart generalist would read it because it shows model-based control, with no training or retraining, handling real objects even when people push or rotate them.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sign error in coupled force-motion model (Eqs. 22-23): Kcoup should subtract, not add, the coupling term; plus sign violates free-object equilibrium.","rationale":"Good-faith reading: the paper is a credible systems contribution with real hardware experiments, reproducible code, and candid limitations. The high-level contact-implicit MPC with warm-start and numerical differentiation is plausible for 10 Hz replanning, and comparisons against MJPC (CEM/iLQG) support the claim that the integrated system is competitive with existing model-based methods. Credit is due for the real-world demos and the admission of low open-box success rates. However, the central low-level modeling contribution—the coupled force-motion model—contains a sign error in Eqs. (22)-(23). The reader's weakest assumption (constant wrench, no inertia) is a modeling concern; the sign error is stronger because it holds even under those assumptions. It is an internal inconsistency that can be settled by re-derivation and a one-dimensional sanity check. If confirmed, the paper's explanation for why the tactile-feedback controller improves robustness is incorrect, even though the empirical system may still work through feedback. Because the error is correctable and the empirical evidence is substantial, I do not move the verdict to REJECT; conditional acceptance should explicitly require fixing the sign and re-validating the coupling term, e.g., by an ablation with the corrected model. The reader's conditions on success-rate reporting and baseline fairness remain valid. My agreement with the reader is partial: we both target the low-level force-motion model, but the specific faulty step is the implicit-function-theorem sign, not only the quasi-static assumptions.","tokens_in":28874,"tokens_out":13653,"duration_ms":130428,"concrete_test":"Verify the sign by re-deriving ∂xu/∂Pd from Eq. (21) with the implicit function theorem and check the 1D two-finger example: with ¯K = kI, Go = [1 1], the correct free-object response to δPd = [δ;δ] is δΛext = 0, while the paper's Eq. (23) gives δΛext = [1.5kδ+0.5kδ; 0.5kδ+1.5kδ] = 2kδ[1;1]. If the re-derivation yields the plus sign and the 1D force change is nonzero, the sign error is confirmed; then correct Eq. (23) to use a minus sign and re-run the Rotate Sphere simulations to verify the coupling term's contribution changes as expected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing flaw is an algebraic sign error in the derivation of the coupled stiffness Kcoup in Sec. 6.3. Starting from δwext = Go δΛext = Go ¯K(δPd − δPc) = 0 and δPc = Go^T δxu (Eqs. 20-21), the implicit function theorem applied to F(δxu, δPd) = Go ¯K δPd − Go ¯K Go^T δxu gives ∂xu/∂Pd = −F_x^{-1} F_Pd = +(Go ¯K Go^T)^{-1} Go ¯K. Equation (22) states the negative of this. Consequently, Eq. (23) should read Kcoup = ¯K − ¯K Go^T(Go ¯K Go^T)^{-1} Go ¯K (a Schur complement), not ¯K + ¯K Go^T(...)^{-1} Go ¯K. The sign is physically meaningful: for a free object grasped by two equal-stiffness contacts, displacing both desired contact points by the same δ should move the object rigidly and leave contact forces unchanged (δΛext = 0); the paper's plus-sign Kcoup predicts δΛext = 2kδ in this 1D example. Since Kcoup enters the low-level dynamics (24), the claimed 'coupling effect' correction has the wrong sign and the model does not capture multi-contact coupling as described. This is an internal mathematical inconsistency, not merely a modeling assumption, and it directly affects the central low-level contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical, model-based framework for in-hand manipulation in which fingers make and break contact. At the high level, a contact-implicit MPC planner based on smoothed quasi-dynamic contact dynamics generates finger motion and contact-force references at 10 Hz; at the low level, an MPC-based hybrid force-motion controller with tactile feedback tracks these references at 30 Hz. The low-level model includes a claimed multi-contact coupling correction. The paper reports simulation comparisons on a rotate-sphere task, robustness studies under noise and model error, and five real-world manipulation tasks, plus a grasping experiment. The central claims are that the integrated motion-contact planning and tracking improves accuracy and robustness relative to existing model-based methods and that the system works without task-specific learning.","tokens_in":29229,"tokens_out":5813,"duration_ms":58587,"significance":"If the claims hold, this is a valuable demonstration that contact-rich in-hand manipulation with regrasping can be performed without task-specific training, and the two-level architecture is a sensible way to combine real-time contact-implicit planning with tactile-feedback tracking. The paper's strengths include the breadth of evaluation, the 100-trial simulation statistics, the use of two different hands, and the availability of code and videos. The main theoretical contribution is the coupled force-motion model in Sec. 6.3, so the correctness of that derivation is decisive for the paper's contribution.","major_comments":[{"comment":"There is a sign error in the derivation of the coupled stiffness. Starting from δwext = GoKbar(δPd − Go^T δxu) = 0, the implicit function theorem gives ∂xu/∂Pd = (GoKbarGo^T)^{-1}GoKbar, with a positive sign, because ∂(δwext)/∂xu = −GoKbarGo^T. Equation (22) states the negative of this result. Consequently Eq. (23) should be a Schur complement, Kcoup = Kbar − KbarGo^T(GoKbarGo^T)^{-1}GoKbar, not Kbar + KbarGo^T(...)GoKbar. The sign is physically meaningful: in a 1D example with two equal-stiffness contacts (Go = [1,1], Kbar = diag(k,k)), displacing both desired contact points by the same δ should move the object rigidly and leave contact forces unchanged, but the printed plus-sign Kcoup predicts δΛext = [2kδ;2kδ]. Since Kcoup enters the low-level dynamics in Eq. (24), the claimed coupling correction is wrong as stated and the low-level model must be re-derived and re-validated.","section":"Sec. 6.3, Eqs. (22)-(23)"},{"comment":"The 'openloop' baseline is not a faithful implementation of Pang et al. (2023): the footnote in Sec. 7.3.2 states that the additional trajectory-optimization stage with a smaller time step is omitted. As implemented, this baseline is an ablation of the proposed framework without low-level tracking, not an external model-based method. The claim that the approach 'outperforms existing model-based methods' should therefore be either supported by a full reproduction of Pang et al. or explicitly reworded, and Table 4 should distinguish the ablation from external baselines.","section":"Sec. 7.3.2 and Table 4"},{"comment":"The open-box task has a success rate acknowledged to fluctuate between 20% and 50% for different boxes, and only two successful trials per box are plotted. The abstract and conclusion claim that the method 'successfully completes five challenging tasks' and is 'robust' overstates the evidence as reported. Please report all trials, the failure modes, and a full success-rate table, and qualify the robustness claim accordingly.","section":"Sec. 8.2.5 and Fig. 20"},{"comment":"The coupled force-motion model is derived under the assumptions wext = Const and quasi-dynamic first-order dynamics, yet the target tasks involve fingers making and breaking contacts and external disturbances. These conditions can produce object acceleration and time-varying external wrenches, so the validity of Eq. (24) for the tested scenarios is not established. Please provide a concrete check, e.g., comparing force predictions of the Kcoup model against a full rigid-body simulation on a representative task, to show the assumption is adequate for control.","section":"Sec. 6.3 and Sec. 3.1"}],"minor_comments":[{"comment":"The numerical-differentiation step size ∆ is never specified; please give its value and comment on its effect on gradient accuracy.","section":"Sec. 5.2.2, Eq. (10)"},{"comment":"The reference 'loco 3d (2025)' is a GitHub issue; please cite a stable archival source or remove the citation.","section":"Sec. 7.2.1"},{"comment":"The axis labels in Fig. 9 are only 'iter', 'yaw', and 'cost'; please add units and a fuller caption so the heatmaps can be interpreted.","section":"Fig. 9"},{"comment":"The sentence explaining that the openloop baseline 'exhibits the lowest standard deviation in task error' due to 'reduced accidental contacts' reads as contradictory and should be clarified.","section":"Sec. 7.3.2"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Sec. 6.3 is load-bearing but local and fixable; I would not reject on it alone. If the authors correct Eqs. (22)-(23), re-validate the low-level model, and address the open-loop baseline and open-box reporting issues, the paper could become acceptable. The scope fit with a robotics journal is good."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a solid systems contribution with five real-world demos, but the stress-test sign error is real. Eq. (22) and (23) in Sec. 6.3 have the wrong sign for the coupling term. From δwext = 0 and δPc = Go^T δxu, the implicit function theorem gives ∂xu/∂Pd = +(Go \\bar K Go^T)^{-1} Go \\bar K, not the negative. So Kcoup should be a Schur complement with a minus, not a plus. This matters because Kcoup enters the low-level dynamics (24) as the claimed multi-contact coupling correction. As written, the model predicts positive force coupling when it should be zero for rigid free-object motion. So the central low-level contribution is mathematically wrong as stated.\n\nThat said, the paper has real value. The high-level contact-implicit MPC with numerical differentiation for contact Hessians is a practical improvement over Pang et al., and the warm-start/shifting tricks are sensible. The automatic weighting-matrix design from planned contact normals is a nice touch. The simulation comparisons are well specified, and the real-world experiments on four LEAP Hand tasks (plus two Allegro sim tasks) are honest and detailed — they even report the open-box success rate at 20-50%, which is candid.\n\nSoft spots beyond the sign error: the openloop baseline in Sec. 7.3.2 is explicitly missing the trajectory optimization step from Pang et al., so it is a weak comparator; the claim that the method 'outperforms existing model-based methods' rests partly on that weakened baseline. The real-world open-box task has a low success rate, and the text plots only successful trials, though it does disclose the range. The low-level model's quasi-static constant-wrench assumption is a further idealization, and the sign error compounds the concern.\n\nBottom line: this deserves a serious referee. The sign error should be caught and fixed (or convincingly argued as a typo that does not affect implementation), and the baseline comparisons need tightening. If the sign is fixed, the paper is a solid model-based in-hand manipulation advance. I'd send it out with a request for a focused check of Sec. 6.3.","headline":"A genuinely integrated planning-tracking system, but the coupled force-motion model has a sign error that invalidates the claimed coupling correction.","tokens_in":29753,"tokens_out":3670,"would_cite":false,"duration_ms":34306,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A purely model-based control hierarchy lets a robotic hand reorient, slide, and open objects through contact-rich in-hand manipulation with no learning or task-specific training.","keywords":["in-hand manipulation","multifingered hands","dexterous manipulation","contact-implicit model predictive control","tactile feedback","hybrid force-motion control","quasi-dynamic contact model","integrated planning and control"],"falsifier":"A decisive check is a high-acceleration disturbance test: grasp an instrumented object with the same three-finger setup, apply a sharp impulse that makes the object accelerate measurably, and compare the low-level MPC's contact-force predictions from the coupled stiffness $K_{\\mathrm{coup}}$ against the tactile readings; systematic divergence under acceleration would show that the coupling model, and the robustness attributed to it, does not hold where inertia matters. The quiet version is a simulation sweep with heavier or faster objects, doubling the object's mass or increasing commanded rotation speed until force-tracking error exceeds the paper's reported levels.","tokens_in":28665,"feed_emoji":"🖐️","tokens_out":11096,"duration_ms":98825,"temperature":0.7,"pith_summary":"The paper sets out to prove that contact-rich in-hand manipulation — where fingers deliberately make and break contact to turn, slide, or open an object — can be performed robustly by a purely model-based control hierarchy, without reinforcement learning or task-specific training. The system couples a high-level contact-implicit model predictive controller that jointly plans finger motions, contact locations, and contact forces through a smoothed, quasi-dynamic contact model, with a low-level MPC that tracks both motion and force references using live tactile feedback through a hand-only force-motion model. The authors' key argument is that the low level exists to absorb precisely the modeling errors that defeat planning-only approaches, in particular the force-at-a-distance artifact of smoothed contacts, which leaves open-loop execution with insufficient contact force. If the claim holds, model-based control becomes a practical alternative to trained policies for new in-hand tasks: the user changes a few hyperparameters, and the same machinery deploys on different hands and objects. The paper reports completion of five real-world tasks under deliberate human disturbances and simulation benchmarks against open-loop trajectory execution and sampling-based MPC.","feed_headline":"Model-based robot hands regrasp objects with no training","feed_subtitle":"A 10 Hz contact-implicit planner plus 30 Hz touch tracking completes five real-world tasks.","key_machinery":"The argument rides on two models. The high-level planner uses the smoothed quasi-dynamic contact (CQDC) dynamics: contact forces emerge as dual variables of a convex optimization with a barrier smoothing term, and the solver is control-limited differential dynamic programming, warm-started from the previous solution and differentiated through numerically computed contact Jacobians; the smoothing parameter $\\kappa$ is the knob that trades nonphysical force-at-a-distance artifacts against the exploration that lets fingers break and re-establish contacts. The low-level controller rests on a force-motion model of the hand alone, whose central identity is the coupled stiffness $K_{\\mathrm{coup}} = \\bar{K} + \\bar{K} G_o^\\top (G_o \\bar{K} G_o^\\top)^{-1} G_o \\bar{K}$, derived by requiring the resultant wrench on the object to stay constant under quasi-static motion; it maps desired fingertip displacements into contact forces across all contacts simultaneously and defines the MPC dynamics $\\dot{s} = g(s,u)$ on the state $[q; q_d; \\Lambda_{\\mathrm{ext}}]$. The weighting matrices derived from planned contact normals — force control along the normal, motion control in the tangent plane — decide which subspace each active contact tracks and turn the same MPC into hybrid force-motion control or plain joint-space control by a switch.","core_discovery":"The central claim is that robustness in contact-rich manipulation decomposes into two jobs that can be assigned to two control layers running at different rates. The high-level layer replans at about 10 Hz through a smoothed contact-implicit dynamics model, generating coarse finger motions and, along with them, reference contact forces and normal directions; this is what recovers from external disturbances. The low-level layer runs at 30 Hz and jointly tracks those references with an MPC whose dynamics describe only the hand: joint positions, commanded positions, and contact forces evolve under a coupled force-motion model. The coupling is the paper's main modeling contribution: the object's free motion links every finger's contact force to every other finger's motion, summarized by the coupled stiffness identity $K_{\\mathrm{coup}} = \\bar{K} + \\bar{K}G_o^\\top(G_o\\bar{K}G_o^\\top)^{-1}G_o\\bar{K}$, so the controller knows that pushing with one finger changes forces elsewhere. With force tracked along the planned contact normal and motion tracked in the tangent plane, the paper reports that real hardware completes door opening, card rotation, board sliding, box opening, and grasping — several under human-imposed disturbances — and that in simulation the method beats open-loop execution and sampling-based MPC on success rate, precision, and smoothness.","pith_inferences":["The same division of labor could transfer to other contact-rich domains, such as legged locomotion: a high-level contact-implicit planner sets foothold forces while a low-level force-tracking layer absorbs terrain-model error, replacing the need for either fast dynamics inversion or heavy domain randomization.","Because the coupled stiffness depends only on grasp geometry and estimated stiffness, it could be estimated or adapted online from tactile and state measurements, turning the constant-wrench quasi-static premise into a continuously checked condition rather than a fixed assumption.","A natural hybrid extension is to keep the 10 Hz contact-implicit planner and replace the analytic low-level stiffness model with a learned residual that predicts force-tracking error; the paper's own reported benchmarks, including the open-box failure rates, would be the baseline for judging whether the residual adds anything.","The authors' stated limit on fast, dynamic tasks points to the next step for the same architecture: closed-form or learned contact dynamics that run faster than the convex contact solver would extend the framework from quasi-static to genuinely dynamic in-hand manipulation."],"forward_implications":["New tasks become configuration changes: the user edits the desired object motion, grasp pose, and coarse models, and the same unmodified pipeline applies — shown across two different hand designs and seven simulated and five physical task types without retraining.","Human-imposed disturbances are absorbed by the 10 Hz planner rather than by fast feedback alone; the card, board, and door trials show errors shrinking back to target after each intervention.","Adding explicit contact-force tracking to motion tracking is what separates the method from planning-only baselines: in the Rotate Sphere benchmark, success rises from 14/100 (open-loop) to 100/100, with far lower joint acceleration.","The smoothing coefficient $\\kappa$ is a genuine design lever: too little smoothing halts finger gaiting, too much introduces nonphysical forces — and the low-level force tracking exists specifically to absorb that artifact.","The 20–50% success rate on the real open-box trials marks the method's practical ceiling: when the true contact geometry deviates strongly from the modeled geometry, model fidelity, not control frequency, limits performance."],"supporting_citations":[{"why":"Supplies the smoothed quasi-dynamic contact-implicit dynamics model (CQDC) that the high-level planner differentiates and optimizes through.","marker":"Pang et al. (2023)"},{"why":"Supplies the control-limited DDP algorithm used to solve the high-level optimal control problem in real time.","marker":"Tassa et al. (2014)"},{"why":"Supplies the single-contact force-motion model and the MPC-based hybrid force-motion control approach that the low level extends with multi-contact coupling.","marker":"Gold et al. (2023)"},{"why":"The paper's prior conference version, from which the high-level motion-contact planning algorithm is refined and extended.","marker":"Jiang et al. (2024)"},{"why":"Supplies the predictive-sampling (CEM) and iLQG MPC baselines that the simulation comparison is measured against.","marker":"Howell et al. (2022)"},{"why":"Supplies the BODex grasp synthesis used to generate the 100 grasp poses for the static force-tracking evaluation.","marker":"Chen et al. (2024)"},{"why":"Supplies the motion-tracking-only baseline and finger-motion planner for the in-grasp object movement experiment.","marker":"Yu et al. (2025)"},{"why":"Supplies the Tac3D tactile sensors that provide the 30 Hz contact-force and normal measurements the low-level controller feeds on.","marker":"Zhang et al. (2022)"}],"fun_headline_variants":["No-training robot hands regrasp via dual-rate control","Model-based regrasping without training data: dual-rate MPC","Robust in-hand regrasp with real-time contact-aware planning","Two-speed controller enables training-free in-hand regrasping","Contact-aware MPC regrasps objects with no learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The low-level force-motion model assumes the object is quasi-static — negligible inertia, rigid point contacts that move with the object, and a constant resultant wrench during the derivation of the coupled stiffness — so its predictions degrade whenever the object accelerates, the contacts deform, or the contact points do not move rigidly with the object, leaving only the high-level replanner as backstop.","fun_headline_variants_meta":{"raw":{"variants":["No-training robot hands regrasp via dual-rate control","Model-based regrasping without training data: dual-rate MPC","Robust in-hand regrasp with real-time contact-aware planning","Two-speed controller enables training-free in-hand regrasping","Contact-aware MPC regrasps objects with no learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000301,"raw_usage":{"total_tokens":1801,"prompt_tokens":1076,"completion_tokens":725,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":642}},"tokens_in":692,"tokens_out":725,"duration_ms":7766,"temperature":1.0,"reasoning_tokens":642,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:15:19.908919+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check is a high-acceleration disturbance test: grasp an instrumented object with the same three-finger setup, apply a sharp impulse that makes the object accelerate measurably, and compare the low-level MPC's contact-force predictions from the coupled stiffness $K_{\\mathrm{coup}}$ against the tactile readings; systematic divergence under acceleration would show that the coupling model, and the robustness attributed to it, does not hold where inertia matters. The quiet version is a simulation sweep with heavier or faster objects, doubling the object's mass or increasing commanded rotation speed until force-tracking error exceeds the paper's reported levels.","supporting_citations":[],"review_version":1}