{"id":"07bdd4fd-dbed-4cf4-8be0-df47ea9f5617","arxiv_id":"2603.09956","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A physics-aware motion-retargeting pipeline that uses ground-reaction-force-derived heel-toe contacts produces dynamically feasible humanoid references and improves downstream imitation learning.","lead":"KDMR retargets human motion-capture and force-plate data into robot reference motions that obey physics constraints, eliminating foot sliding and ground penetration. The authors report that imitation-learning policies trained on these references converge faster and more stably than on purely kinematic retargeting.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The trapezoidal integration constraints (Eqs. 10–11) are written incorrectly (difference instead of sum), so the formulated NLP does not enforce the claimed rigid-body dynamics; this undermines the central dynamic-feasibility claim independent of contact-state assumptions.","rationale":"The reader's weakest assumption focuses on the steady-state contact decomposition, which is plausible and explicitly acknowledged in the limitation section. However, I find the erroneous integration equations in Section III-E more load-bearing because they attack the core mechanism that distinguishes KDMR from kinematic retargeting: the enforcement of dynamics. Even if the contact labels are perfectly correct, the formulated NLP does not enforce the intended relationship between q_t, qdot_t, and qddot_t as written. This makes the central claim of 'dynamically viable' trajectories unsupported from the stated method. The reader did note an 'apparent error in the discrete integration equations' in the rationale, but did not make it the primary concern; hence partial agreement. The paper uses standard tools (CasADi, Pinocchio) and the overall framework is reasonable, so the issue is likely fixable by correcting the equations or by releasing code that shows the correct implementation. That is why I do not move the verdict to REJECT: the concern is addressable and may be a typo. However, it reinforces the need for the conditional verdict: the authors must correct this inconsistency and provide a numerical validation that the optimized trajectories satisfy the continuous dynamics, e.g., via forward simulation. The concrete test proposed would settle whether the published equations are load-bearing or merely a typographical slip. Until then, the central claim cannot be fully trusted.","tokens_in":10919,"tokens_out":5135,"duration_ms":46410,"concrete_test":"Re-run the dynamic optimization with the corrected trapezoidal constraints q_{t+1}=q_t + Δt/2 (qdot_t + qdot_{t+1}) and qdot_{t+1}=qdot_t + Δt/2 (qddot_t + qddot_{t+1}) on the same dataset, then forward-simulate the optimized trajectory with a high-accuracy integrator and report the residual dynamics error. If the residual is nonzero under the corrected constraints, the published equations are not what was implemented; if the trajectory/GRF results change materially, the reported dynamic-feasibility improvement is an artifact of the erroneous constraint.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that KDMR enforces rigid-body dynamics and hence yields dynamically viable references (Abstract, §III-E). The discrete-time constraints that are supposed to enforce this are Eqs. (10)–(11), described as a 'trapezoidal integration law':\nq_{t+1} = q_t + Δt/2 [qdot_{t+1} − qdot_t],\nqdot_{t+1} = qdot_t + Δt/2 [qddot_{t+1} − qddot_t].\nThe standard trapezoidal rule requires the sum of the endpoint velocities/accelerations, not their difference. As written, these constraints do not relate position to velocity or velocity to acceleration at all: if qdot is constant, q_{t+1}=q_t for all t, so the optimized trajectory would have zero displacement while moving. If the implementation follows the paper, the 'dynamically feasible' trajectory need not satisfy the continuous equations of motion (3)–(5); it only satisfies an erroneous algebraic relation. This is more load-bearing than the steady-state contact assumption in §III-C, because even with perfect contact labels, the dynamics enforcement is suspect. The paper does not release code or provide a numerical check (e.g., forward simulation of the optimized trajectory), so this cannot be dismissed as a mere typo without evidence. The limitation section acknowledges the steady-state GRF assumption but does not mention this inconsistency, so it is unacknowledged.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes KDMR, a kinodynamic motion-retargeting pipeline for humanoid locomotion. Human MoCap marker data are converted to human skeleton poses with OpenSim, then to an initial robot reference via differential IK. A heuristic based on the vertical GRF double-peak profile labels heel/flat/toe contact phases and splits the total GRF into heel and toe components. A dynamic trajectory-optimization stage then adjusts the reference using whole-body rigid-body dynamics, contact and friction constraints, and a trapezoidal integration law, producing a dynamically feasible multi-contact reference. The method is evaluated on Unitree G1 with a walking dataset, comparing against the kinematic GMR baseline, and the retargeted references are used to train BeyondMimic policies to assess downstream sample efficiency.","tokens_in":11301,"tokens_out":4262,"duration_ms":42177,"significance":"If the formulation and experiments are correct, the paper addresses an important gap in humanoid motion retargeting: purely kinematic references commonly contain foot sliding, penetration, and dynamically infeasible transitions, which hurt imitation learning. Explicitly embedding heel-toe multi-contact dynamics and GRF-derived contact states into the optimization is a sensible and timely idea. The paper also has concrete strengths: the problem formulation is clearly structured, the downstream policy comparison uses five seeds, and the authors state an intention to open-source the pipeline. However, the current manuscript has a serious error in the integration constraints that directly affects the dynamic-feasibility claim, and the experimental sections rely mostly on qualitative visual comparisons rather than quantitative metrics. For these reasons, the paper needs substantial revision before the claims can be accepted.","major_comments":[{"comment":"The 'trapezoidal integration law' is written as q_{t+1}=q_t + (Δt/2)(qdot_{t+1} - qdot_t) and similarly for velocity. This is not the trapezoidal rule; the correct form uses the sum of the endpoint derivatives: q_{t+1}=q_t + (Δt/2)(qdot_t + qdot_{t+1}). As printed, a constant velocity produces q_{t+1}=q_t, so the constraints do not enforce any relation between position, velocity, and acceleration. Since the central claim of dynamic feasibility in §III-E and the Abstract rests on these constraints, this is a load-bearing issue. Please correct the equations, confirm that the implementation follows the corrected form, and add a numerical consistency check (e.g., forward simulation of the optimized trajectory or reported integration residuals). If this was only a typographical error, state so explicitly.","section":"IV-B, Figs. 5-7"},{"comment":"The experimental evaluation is qualitative. Fig. 5 and Fig. 6a report time traces but give no quantitative metrics for foot sliding, ground penetration, foot-height error, or smoothness; Fig. 7 shows mean reward curves with a shaded region but no numerical values, effect sizes, or statistical tests. The Abstract claims KDMR 'significantly outperforms' GMR. Such claims need numbers: e.g., mean ± std of penetration depth, foot sliding distance, reward after a fixed number of steps, and confidence intervals over the five seeds. Please add a quantitative summary table for the artifact metrics and the policy comparison.","section":"IV-B, Fig. 6b"},{"comment":"The GRF-tracking evaluation is partly circular. The heel/toe contact windows used in the optimization (§III-C) are derived from the same vertical GRF source that is used as the 'Source' curves in Fig. 6b, and the total force is split into heel and toe components using the same heuristic. Matching that decomposition is therefore partly by construction. Please validate against the raw measured total GRF (and ideally independent foot-pressure data), or at minimum quantify and discuss how much of the agreement is attributable to the contact-state labeling heuristic.","section":"IV-C"},{"comment":"The downstream comparison includes only the purely kinematic GMR baseline. While this is appropriate for showing improvement over kinematic retargeting, the paper's framing as a 'kinodynamic' contribution would be strengthened by a comparison with at least one dynamics-aware retargeting baseline, such as DynaRetarget [24] or a simplified dynamics filter. Without such a comparison, the reader cannot tell whether the reported policy improvements come from the specific multi-contact dynamics formulation or simply from any dynamics-aware correction.","section":"III-E"}],"minor_comments":[{"comment":"In the displayed cost function, the plus signs between the force-smoothing term, torque regularization, and ground-contact term are missing, making the expression ambiguous. Please insert them.","section":"III-E, cost function"},{"comment":"The heel/toe contact-window extraction relies on thresholds such as 'periods of non-zero vertical force' and the mid-stance dip. Please specify the exact criteria, thresholds, and smoothing used, since this directly affects reproducibility and the contact labels fed to the optimizer.","section":"III-C"},{"comment":"The phrase 'mathematically guaranteed, smooth and dynamically feasible reference trajectories' overstates what the optimization delivers: the ground-contact constraint (6) is included as a soft cost, the dynamics are discretized, and the integration constraints are only as correct as Eqs. (10)-(11). Please temper the claim to reflect the actual formulation.","section":"I, contribution 1"},{"comment":"The caption says 'foot pitch (φ_z)' and the text mentions 'foot ankle pitch (θ_4)', but the joint naming is inconsistent. Please define the joint ordering and the meaning of θ_Li once in the text.","section":"IV-B, Fig. 5"},{"comment":"The moving-average filter used for visual clarity is not described. Please state the window length and whether the statistics are computed before or after filtering.","section":"IV-C, Fig. 7"},{"comment":"The limitation section acknowledges the steady-state walking assumption and the need for synchronized motion-force data, but it does not mention the integration-constraint issue or the circularity in the GRF-tracking evaluation. Please extend the limitations accordingly after addressing the major comments.","section":"V"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Eqs. (10)-(11) is likely a typographical error rather than a flaw in the implemented solver, but without code or a numerical residual check the reader cannot verify this. If the authors confirm the corrected equations and provide quantitative experimental metrics, the paper could become acceptable. The scope is appropriate for a robotics venue, and the underlying idea is worthwhile; the current presentation is simply not yet rigorous enough to support the strong claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Jim — quick take on arXiv:2603.09956. The central idea is a genuine step forward: instead of doing IK retargeting then hoping for the best, they add a multi-contact trajectory optimization stage that enforces rigid-body dynamics, contact constraints, and friction cones, and they use the vertical GRF curve to infer heel/toe contact windows. That specific combination for humanoid locomotion is new as far as I know, and the downstream comparison against GMR via BeyondMimic is a fair, independent way to test whether the references help a policy. I'm willing to believe the contact-derived references are better.\n\nBut there's a load-bearing problem the paper doesn't address. Equations (10)–(11), the 'trapezoidal integration law,' are written with a minus sign where the trapezoidal rule requires a plus: q_{t+1} = q_t + (Δt/2)(\\dot q_{t+1} - \\dot q_t). With constant velocity that gives q_{t+1}=q_t. So as written, the constraints do not relate positions to velocities or velocities to accelerations at all. If the actual CasADi implementation matches the paper, then the 'dynamically feasible' claim is not supported — the optimization is only enforcing an erroneous algebraic relation, not the equations of motion. This is not a minor typo that can be dismissed without evidence; it's the exact constraint that makes the method kinodynamic. The paper doesn't release code or provide any forward-simulation check, so we can't tell whether it's a typo or a real bug. Either way it has to be fixed and verified.\n\nBeyond that, the empirical section is weaker than the prose. Figures 5–7 show curves but no quantitative metrics or error bars, so 'significantly outperforms' is doing a lot of work. The only baseline is GMR; there's no comparison to DynaRetarget or to the spatio-temporal quadruped retargeting they cite. And the GRF-tracking evaluation in Fig. 6b compares optimized heel/toe forces against a decomposition generated by the same heuristic that set the contact sequences, so it's partly circular. The downstream RL comparison is not circular, which helps, but it needs numbers.\n\nThe limitations section is honest about the steady-state walking assumption and the reliance on synchronized force plates, and that part reads well. It just doesn't mention the integration equations.\n\nIf the authors fix Eqs. 10–11 (or confirm they're a typo), release code, and add a forward-dynamics verification plus a few numbers, this is a solid contribution worth using. As written, I wouldn't trust the central claim. It does deserve a serious referee, though — not a desk reject — because the pipeline is plausible and the problem is real. I'd send it out with a request for major revision and specifically ask for the corrected equations, code, and a replay of the optimized trajectory in simulation. For your reading group, it's a useful case study in why optimization papers should ship reproducibility checks.","headline":"Promising kinodynamic retargeting pipeline, but as written the dynamics constraints in Eqs. 10–11 are wrong and the empirical evidence is under-quantified.","tokens_in":11744,"tokens_out":3081,"would_cite":false,"duration_ms":28003,"reading_group":"yes","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-stage optimizer that adds whole-body dynamics and ground-reaction-force heel–toe contacts to motion retargeting produces dynamically feasible humanoid references and measurably faster imitation learning.","keywords":["kinodynamic motion retargeting","humanoid locomotion","whole-body trajectory optimization","ground reaction forces","multi-contact","imitation learning","heel-toe roll","contact complementarity"],"falsifier":"Retarget a non-steady-state motion (a turn, a stop, a step up, or running) from a similar dataset, or use GRF data interrupted by overground force-plate boundaries, and check whether KDMR still yields contact-consistent, slip-free trajectories; any failure would show the claim holds only for steady-state walking with continuous synchronized GRF.","tokens_in":10842,"feed_emoji":"🤖","tokens_out":3678,"duration_ms":38345,"temperature":0.7,"pith_summary":"The paper argues that fitting human motion-capture data onto a humanoid robot by kinematics alone creates physically impossible references that slow imitation learning. It introduces KDMR, a kinodynamic retargeting pipeline that starts with a kinematic fit, then refines the trajectory through a whole-body optimization enforcing rigid-body dynamics, contact complementarity, friction-cone bounds, and trapezoidal integration. Crucially, KDMR uses ground reaction force (GRF) data to label heel-only, flat-foot, and toe-only contact windows, reproducing the natural heel-to-toe roll. The authors claim this yields trajectories without foot sliding or ground penetration, and that policies trained on them converge faster and reach higher reward than those trained on a purely kinematic baseline.","feed_headline":"Ground-truth forces speed up humanoid motion learning","feed_subtitle":"KDMR turns MoCap and foot-force data into dynamically consistent references, cutting foot slip and boosting policy convergence.","key_machinery":"The load-bearing mechanism is the heel-to-toe roll contact decomposition drawn from the double-peak vertical GRF waveform: a mid-stance dip marks the flat-foot phase with both heel and toe in contact, and the preceding and succeeding non-zero windows are labeled heel-only and toe-only. These binary contact labels drive a multi-contact whole-body trajectory optimization that combines a kinematic initialization (differential inverse kinematics) with a nonlinear dynamic refinement enforcing floating-base dynamics, holonomic contact constraints, a linearized friction pyramid, and trapezoidal integration, with a cost function that penalizes base and joint tracking error, velocity and force discon","core_discovery":"KDMR establishes that integrating GRF-derived contact patterns and explicit multi-contact dynamics into motion retargeting produces dynamically viable reference trajectories that improve downstream imitation learning. The dynamic optimization jointly solves for joint positions, velocities, torques, and heel/toe contact forces subject to the floating-base rigid-body dynamics, zero-velocity and zero-acceleration constraints at active contacts, friction pyramid bounds, and trapezoidal integration, while a weighted cost tracks the kinematic reference, smooths velocities and forces, and penalizes ground penetration. Compared with a purely kinematic retargeting baseline, KDMR removes foot floating","pith_inferences":["The contact-extraction rule is tied to steady-state walking with a characteristic double-peak GRF; a natural stress test is whether the pipeline works on running, turning, stair ascent, or intermittent overground force-plate data, where the waveform differs—if not, the contribution is scoped to walking.","Rescaling GRF by mass ratio and modeling heel and toe as three-dimensional point contacts could make the method transferable across humanoid platforms, but the point-contact assumption may not capture rolling effects on feet with compliant soles.","The reported sample-efficiency gain suggests that many prior imitation results may have been limited as much by reference trajectory quality as by policy architecture or reward design—an implication worth testing on other benchmark motions."],"forward_implications":["Imitation-learned humanoid policies can be trained on dynamically feasible references, so the reinforcement learning agent no longer spends sample budget compensating for impossible motions like foot penetration or base floating.","The GRF-based contact detector may enable natural heel-to-toe roll on humanoids without manual contact scheduling, improving gait stability and stride efficiency.","The two-stage approach—kinematic warm-start followed by dynamic refinement—is a transferable template for other retargeting tasks where contact sequences are known or can be estimated.","Because the dynamic optimization is motion-agnostic once contact sequences are supplied, the method could extend to arbitrary gaits provided a reliable contact estimator exists."],"fun_headline_variants":["Force-fused retargeting stabilizes humanoid locomotion","GRF-inclusive optimization yields dynamically viable humanoid motion","Multi-contact dynamics improve policy convergence in locomotion","Kinodynamic retargeting cuts foot slip for humanoid imitation","Ground reaction forces drive better humanoid motion references"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The contact labels (heel-only, flat-foot, toe-only) are extracted from a steady-state walking GRF profile; if the human motion is not steady-state walking, or if GRF data is intermittent, the contact sequence fed into the dynamic optimizer is wrong and the entire dynamically feasible retargeting rests on incorrect ground contacts.","fun_headline_variants_meta":{"raw":{"variants":["Force-fused retargeting stabilizes humanoid locomotion","GRF-inclusive optimization yields dynamically viable humanoid motion","Multi-contact dynamics improve policy convergence in locomotion","Kinodynamic retargeting cuts foot slip for humanoid imitation","Ground reaction forces drive better humanoid motion references"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000265,"raw_usage":{"total_tokens":1444,"prompt_tokens":742,"completion_tokens":702,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":625}},"tokens_in":486,"tokens_out":702,"duration_ms":7108,"temperature":1.0,"reasoning_tokens":625,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T00:51:57.814594+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retarget a non-steady-state motion (a turn, a stop, a step up, or running) from a similar dataset, or use GRF data interrupted by overground force-plate boundaries, and check whether KDMR still yields contact-consistent, slip-free trajectories; any failure would show the claim holds only for steady-state walking with continuous synchronized GRF.","supporting_citations":[],"review_version":1}