{"id":"cb3c5a3c-bb71-4889-92dd-45b28fdc8cf6","arxiv_id":"2502.07472","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A geometry-free kinematic trajectory optimizer with closed-loop replanning moves grasped objects through 5x5x5 cm goal spaces with roughly 5 mm average position error, winning the RGMC in-hand manipulation track.","lead":"This paper shows a robot hand can slide a grasped object across a 5-centimeter cube of target positions with roughly 5 mm average error, using simple math-based finger planning rather than learned models or object geometry. It reports the winning solution in the ICRA 2024 in-hand manipulation competition and releases code and videos so the result can be checked and reused.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Contact simplification in §IV-B is load-bearing: rolling is 'fully allowed' but no constraint keeps the contact point on the hemispherical fingertip cap, so the reported 5-mm accuracy may rest on closed-loop replanning and compliance rather than the kinematic model.","rationale":"The paper is a genuine real-robot systems contribution: it has a clear optimization, released code/video, analytical gradients, and an external competition result, so I do not think the central empirical claim is fraudulent or hollow. The most load-bearing condition for the method's stated mechanism is that the kinematic simplification in §IV-B is accurate enough that the planned trajectories correspond to a stable in-grasp motion. The reported open-loop errors (~1 cm) and the Appendix C.3 failure under excessive replanning show that this condition is only approximately satisfied and can fail qualitatively when contacts move to non-spherical finger parts. The paper's own evidence thus makes the contact simplification the least secure premise. I agree with the reader's weakest assumption, and I sharpen it: because the fingertips are hemispherical caps on a rigid base, the relevant failure is not just small contact migration but migration off the spherical surface, which invalidates the point-contact-with-friction model. The paper does not quantify the contact point locations in any reported trajectory, so this condition is currently unverified. The concrete check is feasible with the released code and known object/fingertip geometry and would settle whether the assumption holds in the claimed 5-cm regime. If the check passes, the concern is resolved; if it fails, the accuracy should be attributed primarily to closed-loop replanning and physical compliance, and the paper should qualify the 'constant stable grasp' claim. This does not change the reader's CONDITIONAL verdict—it specifies a decisive condition that should be checked before the mechanism claim is taken at face value.","tokens_in":11,"tokens_out":14354,"duration_ms":324831,"concrete_test":"Using the released planner, generate the 40-waypoint trajectories for the known cylinder in the 5×5×5 cm cube (T=3, Nreplan=4, competition weights). With the known cylinder radius/height and the hemispherical fingertip radius (both specified in the paper and Fig. 2), post-hoc compute, at each time step of each planned trajectory, the point on the fingertip sphere that is closest to the cylinder surface. Record the maximum polar angle of the contact point relative to the fingertip's principal axis. If this angle exceeds the angular extent of the silicone hemisphere for any waypoint, the contact would occur on the non-spherical base, contradicting assumption 3. A cleaner variant: run the identical closed-loop protocol with tactile sensors or high-speed video on the fingertips; if contacts leave the spherical cap during nominal runs, the simplification is violated in the claimed regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest link is the contact simplification in §IV-B: fixing each fingertip's center in the object frame while 'fully allowing rolling' is justified by low surface curvature, but the custom fingertips are only hemispherical (Fig. 2), with a 3-D-printed base behind the silicone layer. The optimizer has no constraint that the rolling-induced contact point remain on the spherical cap. During a 5-cm-cube reach, the object can rotate substantially relative to a fingertip; once the contact migrates past the hemisphere rim or onto the non-spherical base, the point-contact/friction model no longer describes the interaction. Appendix C.3 documents this exact failure ('unmodeled contact occurs between the object and non-spherical parts of the fingers') under excessive replanning. The central claim that a constant stable grasp is maintained is therefore not guaranteed by the formulation; it is an empirical property that may hold only for small rolling angles and for the particular initial grasps and replanning counts used. Because the reported success (5 mm average closed-loop error) includes replanning and compliance, it does not establish that the kinematic trajectories themselves respect the assumed contact model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a trajectory-optimization approach to in-grasp object movement for a multi-fingered hand. The method fixes only the fingertip positions in the object frame while allowing rolling contacts, adds a joint-space smoothness cost, and uses SLSQP with analytical gradients to plan full trajectories. A closed-loop scheme re-plans and re-executes until a target error, a replanning budget, or a time limit is reached, and the pipeline returns the fingers to the initial state between waypoints. The authors report winning the RGMC in-hand manipulation track, with average execution errors around 5 mm over 40 waypoints in a 5x5x5 cm cube on a known cylinder and on novel everyday objects, and they compare their approach with a modified re-implementation of a relaxed-rigidity baseline.","tokens_in":17084,"tokens_out":6766,"duration_ms":69056,"significance":"If the reported results hold, the paper shows that a simple, geometry-free kinematic optimizer can achieve practically useful precision and range for in-hand object movement, without pretraining or object models. The strengths include real-world validation on a physical hand, a competition win with externally recorded waypoints, open-source code, and detailed appendices on analytic gradients and variance. The main contribution over prior kinematic in-grasp work is the relaxation of the rigid-thumb-contact assumption, which plausibly enlarges the reachable space while keeping the optimization tractable. The empirical claims are valuable, but the current evidence for the comparative advantage over existing work is weakened by the modified baseline and by missing statistical detail in the central comparison figures.","major_comments":[{"comment":"The pipeline described in Section IV-D returns the fingers to the initial state along the forward trajectory after each waypoint, and Appendix C.2 confirms that the object falls in the second iteration when this return is omitted. This means the reported 40-waypoint experiments are likely 40 out-and-back trials from the center of the cube rather than a continuous traversal from one waypoint directly to the next. The paper should state explicitly whether the object returns to the initial position between waypoints and, if so, reword claims such as 'continuously reach' and 'long-term robustness' so that they describe repeated departures from and returns to the center. This distinction is important for interpreting the difficulty and the validity of the large-range continuous movement claim.","section":"IV-D and Fig. 4"},{"comment":"The baseline implemented for comparison in Section V-C is not the original algorithm of Sundaralingam and Hermans but a modified re-implementation with three deliberate differences: no in-trajectory object-pose cost, joint velocity/movement limits treated as a soft penalty, and the same hyperparameters as the proposed method. These modifications could disadvantage the baseline, so the comparison in Fig. 8 does not by itself support the contribution claim that the proposed approach achieves a larger object reachable space while ensuring task accuracy compared with existing works. The authors should either use the original baseline implementation, justify that each modification is neutral or favors the baseline, or temper the comparative claim to be against this particular re-implementation.","section":"V-C and Appendix F"},{"comment":"The central comparison between the proposed approach and the baseline in Fig. 8 reports only average errors over 40 waypoints without error bars or repeated-trial statistics. Appendix C.1 provides variance data for the proposed method, but no equivalent variance data are given for the baseline. Since the comparative improvement is a stated contribution, the paper should report repeated trials and error bars for both methods, or at least state clearly that the comparison is based on a single continuous run per condition and therefore should be interpreted as anecdotal.","section":"V-C, Fig. 8"},{"comment":"The contact simplification in Section IV-B fixes each fingertip center in the object frame and fully allows rolling, but it places no constraint that the actual contact point remains on the hemispherical fingertip cap. Appendix C.3 documents that with excessive replanning, unmodeled contact occurs between the object and non-spherical parts of the fingers, leading to significant slippage. Because maintaining a constant stable grasp is a central claim, the paper should provide quantitative evidence that for the reported 5 cm motions with the chosen replanning budgets, the contact points remain on the spherical caps (for example, from tracked fingertip-object contact positions or rolling-angle estimates), or explicitly qualify the constant-grasp guarantee as an empirical property of the closed-loop hardware system rather than a property guaranteed by the optimization formulation.","section":"IV-B and Appendix C.3"}],"minor_comments":[{"comment":"The introduction states that in-grasp manipulation can be achieved 'without altering the contact locations,' but Section IV-B explicitly allows rolling contacts that alter contact locations; the terminology should be made consistent, for example by referring to maintaining contact without breaking and re-making contacts.","section":"Section I"},{"comment":"The 'planned error' metric is the terminal pose error of the trajectory that the optimizer minimizes; presenting it as a task-performance benchmark is somewhat circular. Consider relabeling it as 'planned terminal error' and clarifying that it is an optimizer property, with closed-loop execution error being the actual task metric.","section":"Figs. 6, 7, 8, 10"},{"comment":"The comparison of different Nreplan values in Fig. 7 shows averages over eight corners per iteration without error bars; the variability is visible in the scatter points but should be summarized, especially since the conclusion that Nreplan = 8 degrades accuracy in later iterations is based on these averages alone.","section":"Fig. 7"},{"comment":"The collision constraint Fcollision(Qt) is described only as constraining distances between four critical points on the index and ring fingers; the exact definition of Fcollision and the critical points should be stated in the main text or referenced to a specific appendix location for reproducibility.","section":"Eq. (5)"},{"comment":"The pose-goal experiments in Appendix D report a single run for each of the three goals; the paper should state how many trials were performed and whether the reported position and orientation errors are representative, since the appendix elsewhere emphasizes trial-to-trial variability.","section":"Table III"}],"recommendation":"major_revision","confidential_remarks":"The competition result and the open-source release are strong assets, and the core idea is simple and reproducible. The main reservations are about the fairness of the baseline comparison and the need to clarify what 'continuous reaching' means given the move-back-to-initial-state strategy; both are addressable in revision. I do not see a load-bearing error that would require rejection, but the comparative and statistical claims need strengthening before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth a look if you care about practical in-hand manipulation. What is new: it takes the relaxed-rigidity formulation from Sundaralingam and Hermans and loosens it by fixing only fingertip positions in the object frame rather than rigid thumb-object contact, letting all fingertips roll. That does enlarge the reachable space in the experiments, and the closed-loop replanning scheme is a sensible practical addition. The real-robot results are the core: 40 waypoints in a 5 cm cube with about 5 mm average closed-loop error, repeated across runs and on four novel objects, with code and video released. The competition win is supporting external evidence, and the appendix gives a genuinely honest account of failure modes; the excessive-replanning contact degradation in C.3 is exactly the risk you would worry about.\n\nThe stress-test concern about the contact simplification is fair, and the paper is candid about it. The optimization ignores rolling-induced contact migration, justified by low-curvature surfaces and quasi-static motion, and the custom fingertips are only hemispherical over a 3D-printed base. Appendix C.3 confirms that when replanning pushes the contact past the spherical cap, unmodeled contact with non-spherical finger parts causes slippage. So the 5 mm accuracy is not a property of the kinematic model alone; it is the closed-loop system, including PD compliance and soft fingertips, that makes it work. That does not break the main claim, because the paper claims the approach achieves the accuracy, not that the simplified model predicts it perfectly. But the limitation is load-bearing and should be stated more prominently than a brief assumption in Section IV-B.\n\nMinor soft spots: the main comparison figures lack error bars, though Appendix C.1 partially fixes that; the baseline in Section V-C is a modified re-implementation, disclosed in Appendix F; hyperparameters were tuned on the same benchmark used for the headline numbers; and the 7/9 cm reachable-space results involve manual grasp adjustments and per-goal replanning choices, making those numbers less clean. None of these are fatal, but they should be disclosed in the main text.\n\nBottom line: a solid systems contribution, honestly reported, with reproducible artifacts. It is an extension of an existing framework rather than a new paradigm, and its practical value is the main asset. A serious referee should engage with it. I would send it to review, with a request to move the contact-simplification caveat and the hyperparameter-tuning disclosure into the main text. I would cite it if I were working on geometry-free in-hand manipulation.","headline":"A useful, honest systems paper that extends relaxed-rigidity optimization by allowing all-fingertip rolling; the simplified contact model is a real limitation but the closed-loop empirical results carry the claim.","tokens_in":17670,"tokens_out":3555,"would_cite":true,"duration_ms":31826,"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":"This paper claims that precise, large-range in-grasp object movement can be achieved by a kinematic optimizer that fixes fingertip positions, allows full rolling, and closes the loop with replanning—with no object geometry or training.","keywords":["multi-fingered in-hand manipulation","in-grasp object movement","trajectory optimization","rolling contact","closed-loop replanning","kinematic planning","dexterous manipulation","RGMC"],"falsifier":"Measure fingertip contact migration during a 40-waypoint closed-loop run, for example with a tactile array or a high-speed camera; if the accumulated contact migration exceeds the hemispherical fingertip radius, the fixed-position surrogate no longer predicts the actual motion. Alternatively, run the identical pipeline on a small-radius sphere (about 2 cm radius) with the same replanning cap: if the object drops or average execution error grows well beyond the reported 5 mm, the low-curvature premise is falsified.","tokens_in":16629,"feed_emoji":"🤖","tokens_out":5692,"duration_ms":53975,"temperature":0.7,"pith_summary":"The paper claims that a simple kinematic trajectory optimizer can move an object grasped in a multi-fingered hand through a large range of positions with millimeter-level precision, without any object geometry, pretraining, or contact simulation. The trick is to enforce the stable grasp only by fixing each fingertip's position in the object frame, leaving fingertip orientation free so all fingertips may roll. Closed-loop replanning from the measured current pose absorbs the errors this simplification introduces. If right, this makes precise in-grasp manipulation practical for novel everyday objects on generic dexterous hands.","feed_headline":"No-geometry planner hits 5 mm in-hand accuracy over 5 cm","feed_subtitle":"Champion solution moves known and novel objects through 40 waypoints with no pretraining and no object shapes.","key_machinery":"The load-bearing object is the cost $J = J_{\\text{object}} + J_{\\text{finger}} + J_{\\text{joint}}$, where $J_{\\text{object}}$ penalizes the terminal pose error via the weighted SE(3) distance $d(T_1,T_2,W)$ with error $e=[p_e; r_e]$, $J_{\\text{finger}}$ penalizes changes in each fingertip's pose relative to the object frame with orientation weights near zero, and $J_{\\text{joint}}$ regularizes the joint trajectory. The constant-stable-grasp requirement is replaced by the surrogate 'fingertip positions fixed in the object frame', which ignores rolling-induced contact migration and is justified by an assumption of smooth, low-curvature surfaces and quasi-static motion. Analytic gradients come from perturbing the SE(3) pose, using the left Jacobian of SO(3) and a relative Jacobian between fingertip and object frames; the non-convex problem is solved sequentially with SLSQP. A closed-loop scheme re-plans and re-executes until planned error, replanning count, or time budget is met.","core_discovery":"The paper's central claim is that in-grasp object movement—moving an object's pose using only finger motion under a constant stable grasp—can be solved by a geometry-free kinematic trajectory optimizer. The constant-grasp requirement is reduced to a surrogate: keep each fingertip's position fixed in the object frame, but allow all fingertips to roll. Assigning zero orientation weight in the fingertip cost makes rolling free, and optimizing the object pose as a decision variable frees the thumb from the rigid-contact assumption that limited prior relaxed-rigidity approaches. Closed-loop replanning from the current measured state to the goal compensates for the rolling simplification, and returning the fingers to the initial grasp between waypoints preserves long-horizon robustness. On the RGMC task this yields average execution errors near 5 mm over 40 continuous waypoints inside a 5×5×5 cm cube, on both the known cylinder and novel objects, without using object geometry.","pith_inferences":["Editorial extension: the fixed-fingertip-position surrogate could serve as a cheap warm-start or feasibility filter for contact-implicit or learning-based planners, because it produces joint-limit- and collision-respecting trajectories in seconds rather than the per-step cost of rigorous rolling models.","Editorial extension: since the object cost can carry orientation weights, the formulation implies reachable pose goals with rotation; the paper demonstrates only three such goals, leaving full six-dimensional waypoint sequences as a natural open test.","Editorial extension: the documented failure mode under excessive replanning suggests a closed-loop regulator that monitors contact quality (for instance by tactile sensing) could outperform the fixed replanning cap, an option the authors do not explore.","Editorial extension: the screwdriver results, which required occasional manual grasp resets, hint that the low-curvature assumption is the method's boundary; a systematic sweep over surface curvature would quantify where the surrogate breaks."],"forward_implications":["Allowing rolling at every fingertip, rather than keeping the thumb contact rigid, enlarges the object's reachable space and improves both planned and executed accuracy compared with the rigid-thumb baseline.","Closed-loop replanning roughly halves execution error, from about 10 mm open-loop to about 5 mm, and keeps errors stable across 40 consecutive waypoints when the replanning cap is moderate.","The same pipeline, with no object model, transfers to novel everyday objects—thick cylinder lid, box, presenter remote, and screwdriver—with average errors around 5 mm, even though high curvature and weight increase error.","Accuracy degrades gracefully with distance: sub-millimeter errors at 1 cm cube corners, and roughly 2 cm average error at the extremes of a 9×9×9 cm space, showing the planner exploits the fingers' workspace.","Excessive replanning degrades contact quality, so the competition strategy used a conservative replanning cap in one run and an aggressive cap in the other to balance robustness and precision."],"supporting_citations":[{"why":"Supplies the relaxed-rigidity kinematic trajectory optimization baseline whose rigid-thumb assumption this paper replaces with rolling contacts.","marker":"[5]"},{"why":"Journal version of the same relaxed-rigidity approach, providing the comparison baseline and the initial inspiration for the formulation.","marker":"[6]"},{"why":"Defines the RGMC competition whose in-hand manipulation track sets the task, objects, scoring, and the championship claim.","marker":"[7]"},{"why":"Describes the Leap Hand hardware used in all real-world experiments and in the competition.","marker":"[22]"},{"why":"Recent trajectory optimization with rigorous rolling and geometric constraints, cited to show the complexity (around 8 seconds per MPC step) avoided by the simplified surrogate.","marker":"[25]"},{"why":"The SLSQP sequential quadratic programming implementation used to solve the non-convex trajectory optimization problem.","marker":"[28]"},{"why":"Provides the relative Jacobian expression between fingertip and object frames needed for the analytic gradients of the fingertip cost.","marker":"[30]"},{"why":"The inverse-kinematics solver used to convert the manually defined fingertip positions into the joint-space command for the initial grasp.","marker":"[31]"}],"fun_headline_variants":["Champion in-hand plan: no geometry, no pretraining, 5 mm accuracy over 5 cm","Geometry-free in-hand optimizer wins RGMC, 5 mm precision over 5 cm","No object shapes needed: RGMC champion hits 5 mm in-hand accuracy","Rolling fingers, fixed grasp: champion planner achieves 5 mm over 5 cm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes that ignoring fingertip rolling and contact migration—treating a stable grasp as fixed fingertip positions in the object frame—still produces the planned motion, which holds only if the object surface is smooth and low-curvature and the motion is quasi-static.","fun_headline_variants_meta":{"raw":{"variants":["Champion in-hand plan: no geometry, no pretraining, 5 mm accuracy over 5 cm","Geometry-free in-hand optimizer wins RGMC, 5 mm precision over 5 cm","No object shapes needed: RGMC champion hits 5 mm in-hand accuracy","Rolling fingers, fixed grasp: champion planner achieves 5 mm over 5 cm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000559,"raw_usage":{"total_tokens":2654,"prompt_tokens":936,"completion_tokens":1718,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":1625}},"tokens_in":552,"tokens_out":1718,"duration_ms":11360,"temperature":1.0,"reasoning_tokens":1625,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T12:36:41.182001+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure fingertip contact migration during a 40-waypoint closed-loop run, for example with a tactile array or a high-speed camera; if the accumulated contact migration exceeds the hemispherical fingertip radius, the fixed-position surrogate no longer predicts the actual motion. Alternatively, run the identical pipeline on a small-radius sphere (about 2 cm radius) with the same replanning cap: if the object drops or average execution error grows well beyond the reported 5 mm, the low-curvature premise is falsified.","supporting_citations":[{"cited_title":"Relaxed-rigidity constraints: In- grasp manipulation using purely kinematic trajectory optimization,","cited_arxiv_id":null,"evidence_quote":"Supplies the relaxed-rigidity kinematic trajectory optimization baseline whose rigid-thumb assumption this paper replaces with rolling contacts."},{"cited_title":"Relaxed-rigidity constraints: kinematic trajectory optimization and collision avoidance for in-grasp manipulation,","cited_arxiv_id":null,"evidence_quote":"Journal version of the same relaxed-rigidity approach, providing the comparison baseline and the initial inspiration for the formulation."},{"cited_title":"9th Robotic Grasping and Manipulation Competition","cited_arxiv_id":null,"evidence_quote":"Defines the RGMC competition whose in-hand manipulation track sets the task, objects, scoring, and the championship claim."},{"cited_title":"LEAP Hand: Low-cost, efficient, and anthropomorphic hand for robot learning,","cited_arxiv_id":null,"evidence_quote":"Describes the Leap Hand hardware used in all real-world experiments and in the competition."},{"cited_title":"A software package for sequential quadratic programming,","cited_arxiv_id":null,"evidence_quote":"The SLSQP sequential quadratic programming implementation used to solve the non-convex trajectory optimization problem."},{"cited_title":"A more compact expression of relative jacobian based on individual manipulator jacobians,","cited_arxiv_id":null,"evidence_quote":"Provides the relative Jacobian expression between fingertip and object frames needed for the analytic gradients of the fingertip cost."},{"cited_title":"In-hand following of deformable linear objects using dexterous fingers with tactile sensing,","cited_arxiv_id":null,"evidence_quote":"The inverse-kinematics solver used to convert the manually defined fingertip positions into the joint-space command for the initial grasp."}],"review_version":1}