{"id":"fe41ee92-6d1b-4179-8040-bdac307ff68d","arxiv_id":"2505.19540","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A ZMP-based pendulum-plus-full-body-kinematics model with an MLP warm-start runs whole-body MPC for bipedal walking in under 17 ms and survives pushes in simulation and on the TOCABI humanoid.","lead":"This paper builds a real-time whole-body controller for two-legged robots by mixing a simple pendulum-style model with full-body motion constraints, and uses a small neural network to guess good starting solutions for the optimizer. The result is a controller that recomputes the whole-body plan in under 17 milliseconds, which the authors test on a real humanoid robot under pushes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The constant-COM-height LIPFM in Eqs. (1)-(2) and the full-body COM constraint in Eq. (5) are never reconciled; on the flat-ground push experiments themselves, whole-body recovery motions may violate the assumed fixed z_c, so the MPC plan may not describe the executed centroidal motion.","rationale":"The reader's weakest assumption correctly identifies the constant-COM-height limitation, but frames it as a terrain-scope restriction. I agree with that observation and sharpen it into a potential internal inconsistency: the flat-ground push experiments themselves excite whole-body motions that can change COM height, so the concern is not only about stairs or slopes. The paper never specifies whether h1 fixes X_com,z(q) = z_c or leaves vertical COM unconstrained, and it reports no COM-height data from the experiments. The proposed check directly settles whether the LIPFM assumption and the FKM constraint are mutually consistent in the demonstrated regime. This does not change the overall verdict: the paper remains a plausible engineering contribution, conditional on this consistency check, on reporting total control-loop latency including the WBC, and on the fairness of the CP-MPC baseline. Therefore the reader's CONDITIONAL verdict is retained unchanged.","tokens_in":11155,"tokens_out":12114,"duration_ms":120350,"concrete_test":"From the recorded MuJoCo simulation or real-robot logs used for Fig. 7, compute the executed COM height z_com(t) = \\bar X_com,z(q(t)) and compare it with the constant z_c assumed in Eqs. (1)-(2), and also compare the LIPFM predicted COM (x(t), y(t)) with \\bar X_com,x(q(t)), \\bar X_com,y(q(t)) at the MPC nodes. Report max |z_com(t) - z_c| during nominal walking and after the 30.4 Ns push, along with the penalty residuals of Eq. (5). If the deviation exceeds a few centimeters, or if the LIPFM COM and FKM COM disagree beyond the penalty tolerance, the kino-dynamic model is not self-consistent in the very regime used to support the central claim; balance would then need to be re-attributed to the WBC/feedback loop. If the deviations are negligible, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed kino-dynamic model enables real-time whole-body MPC that maintains balance. The most load-bearing concern is an internal-consistency gap between the two halves of that model. Equations (1)-(2) treat z_c as a fixed parameter in the LIPFM, while Eq. (5) defines the COM from the full-body kinematics, x_com = \\bar X_com(q), which includes a vertical component. The paper never states whether h1 enforces X_com,z(q) = z_c, nor whether it enforces consistency between the LIPFM COM state (x, y, \\dot x, \\dot y) and the COM velocity/acceleration implied by q and \\dot q. If X_com,z(q) is allowed to vary, the MPC can command exactly the whole-body recovery motions shown in Fig. 7 (pelvis pitch, waist pitch, shoulder motion) while Eqs. (1)-(2) assume a constant COM height; the planned ZMP/COM trajectory then does not match the robot's actual centroidal motion. If X_com,z(q) = z_c is instead enforced, the whole-body motions are artificially restricted and the claimed dynamic effects of whole-body motion are not actually captured. Either way, the 17 ms solver latency may be real, but the optimized plan is not shown to be a valid kino-dynamic plan in the demonstrated scenarios. The robot's balance could then be maintained by the WBC and PD feedback rather than by the MPC model. The authors' own conclusion concedes limits on vertical COM motion for stairs and slopes, but the same issue can arise on flat ground whenever the perturbation response changes COM height; the paper provides no measurement of this.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a real-time whole-body model predictive control framework for bipedal locomotion. The authors introduce a 'kino-dynamic' model that couples the linear inverted pendulum plus flywheel model with full-body forward kinematics, replacing contact-wrench formulations with ZMP-based constraints. They also propose a modularized multi-layer perceptron warm-start strategy that predicts an initial guess for the DDP solver at each control cycle, and a ZMP-based whole-body controller. The reported results include latency comparisons against WBD and centroidal-dynamics baselines, warm-start accuracy comparisons, and simulated and real-robot push-recovery experiments on the TOCABI humanoid, with a central quantitative claim of solving a 60-step, 4620-variable, 2400-constraint WB-MPC within 17 ms.","tokens_in":11524,"tokens_out":7871,"duration_ms":69806,"significance":"If the central claims hold, the paper is a useful practical step toward real-time whole-body MPC on torque-controlled humanoids: the proposed model keeps per-iteration latency low and, importantly, low-variance across contact transitions, and the real-robot validation under perturbations is valuable. The explicit reporting of worst-case latencies and the candid statement of the model's vertical-COM limitations are also strengths. The significance is tempered by an unresolved consistency question between the LIPFM's constant-COM-height assumption and the full-body kinematic constraints, and by a robustness comparison that removes the footstep-adjustment mechanism from the CP-MPC baseline. These issues affect the scope of the central claim more than the measured latency numbers.","major_comments":[{"comment":"The manuscript does not reconcile the constant-COM-height assumption of the LIPFM with the full-body COM map \\bar X_com(q) that appears in the h1 constraints. Because the MPC state contains both the generalized coordinates q and the LIPFM COM variables (x, y, \\dot x, \\dot y), the formulation must state whether h1 enforces X_com,z(q) = z_c, and whether it ties the LIPFM COM position/velocity to the COM position/velocity implied by (q, \\dot q). Neither is specified. If the vertical COM coordinate is left free, the whole-body recovery motions shown in Fig. 7 (pelvis pitch, waist pitch, shoulder motion) can change z_c while Eqs. (1)-(2) treat z_c as fixed, so the planned ZMP/COM trajectory may not describe the executed centroidal motion. If X_com,z(q) = z_c is imposed instead, the dynamic effects of whole-body motion are artificially restricted. In either case the 17 ms solver latency does not by itself establish that the optimized plan is physically valid in the demonstrated scenarios. The conclusion's caveat about stairs and slopes is the same vertical-COM issue restricted to those terrains; it can also arise on flat ground under perturbation. Please state the exact form of h1 and report the vertical COM trajectory from the Fig. 7 experiment.","section":"Section II-A/II-B, Eqs. (1)-(2), (5), and (9)"},{"comment":"The robustness comparison against CP-MPC removes the footstep-adjustment strategy of CP-MPC, despite that strategy being CP-MPC's primary mechanism for responding to perturbations. As a result, 'our proposed WB-MPC outperforms CP-MPC' is not a comparison against CP-MPC as designed. Please either include the footstep strategy in the baseline and compare under the same footstep constraints, or rephrase the conclusion to state that the proposed controller outperforms a CP-MPC variant without stepping. The acknowledgment that an ideal-real-time WBD baseline withstands larger perturbations is helpful, but it implies the robustness advantage over WBD rests on that baseline's non-real-time latency; the summary comparison should say so explicitly.","section":"Section V-A and Fig. 6(a)"},{"comment":"The modularized MLP is trained on trajectories generated by the same WB-MPC solver that it warm-starts, so the reported reductions in DDP iterations and residual cost are partly a measure of how well the regressor reproduces solver outputs within its training distribution. The paper mentions stable walking for 'unseen states' and different step lengths, but it does not quantify generalization to perturbation magnitudes or directions that were absent from the training set. Please state the perturbation sampling distribution used to create the motion datasets and report held-out results (for example, residual cost and average iterations for push magnitudes outside the sampled range) so that the warm-start claim is tested beyond interpolation of the solver's own outputs.","section":"Section III and Fig. 5"}],"minor_comments":[{"comment":"The caption contains a typo: 'referes' should be 'refers'. Please also report the number of trials underlying the mean and maximum latencies in Fig. 6(b).","section":"Fig. 6"},{"comment":"The dynamics g in Eq. (8) is never written out. In particular, the evolution of the CAM components h_x and h_y is not stated; please provide the explicit continuous-time or discretized dynamics, including \\dot h_x = \\tau_x and \\dot h_y = \\tau_y (or equivalent), so the LIPFM state evolution is well-defined.","section":"Eq. (8)"},{"comment":"The numerical values of the cost weights W_p, W_pl, W_cp, W_e, W_u and the WBC weights W_c, W_q, W_F are not reported, and the RBF/PCA hyperparameters are described only as 'fine-tuned via line search'. Without these values, the comparative experiments are difficult to reproduce.","section":"Section II-B cost weights"},{"comment":"Notation is inconsistent: Eq. (4) writes the end-effector map as \\bar X_e while the text refers to \\bar X_c; please unify the notation and define T_{e,d} more precisely as a desired SE(3) placement.","section":"Eqs. (4)-(5)"},{"comment":"The R-squared values 0.99975 and 0.99953 are cited for PCA and RBF, but it is not stated whether they measure reconstruction accuracy in the reduced space or the original trajectory space; please clarify.","section":"Table III"},{"comment":"The 12 real-robot trials are summarized only by maxima and averages; reporting per-condition mean and standard deviation of the achieved DDP latency and iteration count would strengthen the claim.","section":"Section V-B"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a real engineering achievement—whole-body MPC on a torque-controlled humanoid at 50 Hz with peak latency under 17 ms, including during pushes—but the physical consistency of the model is not fully nailed down, and the comparison baselines are somewhat cherry-picked.\n\nWhat's genuinely new: the LIPFM-plus-FKM formulation that uses ZMP instead of contact wrenches, avoiding the latency spikes during contact transitions; and the modularized MLP warm-start with PCA/RBF compression, which cuts prediction latency from ~500 ms to ~0.3 ms while keeping accuracy close to GPR. The DDP solver stats are well measured: average 2-3 iterations, low variance, and real robot data supporting the latency claims. That part holds up.\n\nThe soft spots are real. First, the constant-COM-height LIPFM (Eqs. 1-2) and the full-body COM constraint (Eq. 5) are never reconciled. The paper doesn't state whether the MPC enforces X_com,z(q)=z_c. If it does, the upper-body recovery motions shown in Fig. 7 are artificially constrained; if it doesn't, the LIPFM is being used outside its validity. Either way the optimized plan may not describe the executed centroidal motion, and the robustness could be coming from the WBC and PD loop. This is not just a stairs problem; it can matter on flat ground during large push recovery. The authors' own conclusion concedes vertical COM limitations but doesn't address the flat-ground case.\n\nSecond, the comparison handicaps CP-MPC by removing its footstep adjustment, and the authors admit that an ideal-real-time WBD baseline handles larger perturbations. So the contribution is computational speed, not improved robustness over an unconstrained WBD.\n\nThird, no code, data, or weight values are released, which hurts reproducibility.\n\nOn the other hand, the circularity of training the warm-start on the same solver is partially mitigated by benchmarking against cold start and other learned predictors. The latency evidence is concrete. For a robotics audience, this is a solid, significant step: it makes real-time whole-body MPC practical, which is a genuine bottleneck. The model inconsistency should be addressed—at minimum, report the vertical COM error during the experiments and quantify how much the MPC constraint is violated.\n\nWho should read it: anyone working on humanoid MPC or warm-starting DDP. It deserves a serious referee even with these caveats; the core claim is important and mostly supported. My recommendation: send it to peer review, but require the authors to clarify the COM-height constraint and release hyperparameters, and rerun the CP-MPC comparison without handicapping the stepping.","headline":"Real whole-body MPC at 50 Hz on a humanoid is a genuine engineering win, but the model's COM-height consistency and some baseline choices need scrutiny before this is treated as settled.","tokens_in":12076,"tokens_out":2914,"would_cite":true,"duration_ms":24491,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A kino-dynamic model that merges the LIPFM with full-body kinematics and a modularized MLP warm-start lets whole-body MPC for bipedal walking run within a 20 ms control cycle, solving 60-step problems with 4620 variables in under 17 ms.","keywords":["whole-body model predictive control","kino-dynamic model","linear inverted pendulum plus flywheel","ZMP-based whole-body control","modularized MLP warm-start","bipedal locomotion","real-time control","humanoid robot"],"falsifier":"Run the framework with the same solver settings on a walking task with a vertical step large enough to force the center of mass to rise or fall by several centimeters, and check whether the robot stays balanced and the solve stays under 20 ms; a failure would show that the constant-COM-height kino-dynamic model, not the warm-start, is the limiting assumption.","tokens_in":10943,"feed_emoji":"🤖","tokens_out":8139,"duration_ms":73232,"temperature":0.7,"pith_summary":"This paper claims that whole-body model predictive control (WB-MPC) for bipedal walking can be executed inside a 20 ms control cycle when the planner uses a kino-dynamic model built from the linear inverted pendulum plus flywheel (LIPFM) and full-body kinematics, with the zero-moment point (ZMP) replacing contact wrenches as the contact representation. The accompanying warm-start strategy uses a modularized multi-layer perceptron to predict a good initial guess for each control cycle, reducing warm-start latency from hundreds of milliseconds to under a millisecond. The authors report solving a 60-time-step, 4620-variable, 2400-constraint WB-MPC in at most about 17 ms, including under external pushes, and validate the framework in simulation and on the TOCABI humanoid. The paper's own conclusion restricts the claim to flat-ground walking; stairs, steep slopes, and flight phases are outside the model.","feed_headline":"Bipedal whole-body MPC solved in 17 ms","feed_subtitle":"A LIPFM-plus-kinematics model with MLP warm-start keeps a humanoid balanced under pushes in real time.","key_machinery":"The load-bearing object is the kino-dynamic model that merges the LIPFM with full-body kinematics. Equations (1)-(2) give COM acceleration as a linear function of ZMP and the derivative of centroidal angular momentum at constant COM height $z_c$; equation (3) connects joint accelerations to that momentum derivative through the centroidal momentum matrix; and equations (4)-(5) enforce COM and foot placements through forward kinematics. Because the contact quantity is the ZMP rather than distributed contact wrenches, the optimization constraints do not change form when the foot support phase changes, and no inverse dynamics is needed to build the model. The other central piece is the warm-start: a set of small single-hidden-layer MLPs, one for each walking phase and each variable group, maps a low-dimensional input state to RBF/PCA coefficients of a near-optimal state trajectory, which the DDP solver then refines.","core_discovery":"The central claim is that a whole-body MPC for bipedal locomotion can use ZMP-based dynamics instead of contact-wrench whole-body dynamics and still capture the stabilizing effects of whole-body motion. The proposed kino-dynamic model couples the LIPFM equations of the center of mass to the centroidal momentum matrix and forward kinematics, so joint motion, feet placement, and angular momentum enter the optimization without an inverse-dynamics pass. Replacing contact wrenches with ZMP keeps the problem structure unchanged across support phases, which removes the peak latency observed at contact transitions in wrench-based whole-body MPC and lowers the average per-iteration cost by roughly 44 percent in the paper's comparisons. With the modularized MLP warm-start, the solver converges in about two DDP iterations per cycle, and the combined system keeps the robot balanced against pushes in real time; on hardware, it withstands a 30.4 Ns frontal impact and a 15.3 Ns lateral impact.","pith_inferences":["The constant-COM-height assumption suggests the framework is a flat-ground, stance-phase planner; extending it to stairs, slopes, running, or jumping would require replacing or augmenting the LIPFM equations, and the reported latency would likely change.","Because the MLP is trained per walking phase and per variable type, its coverage of unseen gaits is bounded by the training distribution; the paper's successful unseen-state walk is suggestive, but a systematic sweep over step length, speed, and terrain height would map the true generalization region.","The phase-invariant structure of the model implies that the latency advantage over wrench-based whole-body MPC should persist on higher-degree-of-freedom humanoids, though the absolute 17 ms number is tied to the 19-actuator configuration tested."],"forward_implications":["Humanoid controllers can embed whole-body dynamic effects into the MPC layer itself, since the planned trajectories for pelvis, waist, shoulders, and feet are computed reactively rather than by a separate simplified footstep planner.","The reported worst-case latency of about 17 ms means the same 20 ms control loop can close on a torque-controlled humanoid without splitting planning and control into separate rates.","Warm-start latency no longer scales with the size of the motion database, so the training set can include perturbation scenarios and varied initial states without breaking the real-time budget.","ZMP-based whole-body control can be merged with the MPC objective so that impulses at foot landing and ZMP regulation are handled in the same feedback law."],"supporting_citations":[{"why":"Supplies the LIPFM equations that the proposed kino-dynamic model is built on.","marker":"[2]"},{"why":"Defines the memory-of-motion warm-start strategy and the RBF/PCA dimensionality reduction that the modularized MLP adapts.","marker":"[9]"},{"why":"Shows a whole-body MPC with memory of motion on a torque-controlled humanoid, the real-time target this paper aims to achieve.","marker":"[10]"},{"why":"Provides the wrench-based whole-body MPC baseline whose latency and contact-transition behavior are compared against the proposed model.","marker":"[11]"},{"why":"Supplies the penalty method used to handle full-body kinematic constraints inside the MPC.","marker":"[15]"},{"why":"Provides the control-limited feasibility-driven DDP solver used to solve the WB-MPC.","marker":"[16]"},{"why":"Describes the TOCABI humanoid used for the simulations and real-robot experiments.","marker":"[17]"},{"why":"Provides the capture-point MPC baseline used in the perturbation robustness comparison.","marker":"[18]"}],"fun_headline_variants":["Real-time bipedal MPC via ZMP and MLP warm-start","ZMP-based model cuts whole-body MPC cost by 44%","MLP warm-start enables real-time bipedal MPC","Bipedal whole-body MPC robust to pushes with ZMP model","ZMP + MLP makes whole-body MPC real-time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes the center of mass stays at a fixed height, so the real-time planning claim only covers flat-ground walking without flight phases; stairs, steep slopes, running, and jumping are outside what the MPC can represent.","fun_headline_variants_meta":{"raw":{"variants":["Real-time bipedal MPC via ZMP and MLP warm-start","ZMP-based model cuts whole-body MPC cost by 44%","MLP warm-start enables real-time bipedal MPC","Bipedal whole-body MPC robust to pushes with ZMP model","ZMP + MLP makes whole-body MPC real-time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000816,"raw_usage":{"total_tokens":3606,"prompt_tokens":1004,"completion_tokens":2602,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":2514}},"tokens_in":620,"tokens_out":2602,"duration_ms":12813,"temperature":1.0,"reasoning_tokens":2514,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:12:11.400901+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the framework with the same solver settings on a walking task with a vertical step large enough to force the center of mass to rise or fall by several centimeters, and check whether the robot stays balanced and the solve stays under 20 ms; a failure would show that the constant-COM-height kino-dynamic model, not the warm-start, is the limiting assumption.","supporting_citations":[{"cited_title":"Versatile reactive bipedal locomotion planning through hierarchical optimization,","cited_arxiv_id":null,"evidence_quote":"Supplies the LIPFM equations that the proposed kino-dynamic model is built on."},{"cited_title":"Learning How to Walk: Warm-Starting Optimal Control Solver with memory of motion,","cited_arxiv_id":null,"evidence_quote":"Defines the memory-of-motion warm-start strategy and the RBF/PCA dimensionality reduction that the modularized MLP adapts."},{"cited_title":"Whole Body Model Predictive Control with a memory of motion: Experiments on a Torque-Controlled Talos,","cited_arxiv_id":null,"evidence_quote":"Shows a whole-body MPC with memory of motion on a torque-controlled humanoid, the real-time target this paper aims to achieve."},{"cited_title":"Whole-Body Model Predictive Control for Biped Locomotion on a Torque-Controlled Humanoid Robot,","cited_arxiv_id":null,"evidence_quote":"Provides the wrench-based whole-body MPC baseline whose latency and contact-transition behavior are compared against the proposed model."},{"cited_title":"Whole-Body MPC and Online Gait Sequence Generation for Wheeled-Legged Robots,","cited_arxiv_id":null,"evidence_quote":"Supplies the penalty method used to handle full-body kinematic constraints inside the MPC."},{"cited_title":"Crocoddyl: An efficient and versatile framework for multi-contact optimal control,","cited_arxiv_id":null,"evidence_quote":"Provides the control-limited feasibility-driven DDP solver used to solve the WB-MPC."},{"cited_title":"Design of the humanoid robot TOCABI,","cited_arxiv_id":null,"evidence_quote":"Describes the TOCABI humanoid used for the simulations and real-robot experiments."}],"review_version":1}