{"id":"12db4816-40ab-4056-a097-bae59f5b5c1c","arxiv_id":"2504.18698","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A unified MPC using an augmented LIP model with ZMP dynamics enables a simulated Cassie biped to recover from large frontal and lateral pushes by simultaneously adjusting foot placement, step timing, and ankle effort.","lead":"The authors combine foot placement, step timing, and ankle/ZMP control into one model predictive controller on a reduced-order model of walking. They demonstrate in simulation that this combined strategy recovers a Cassie biped from pushes that single-strategy controllers cannot handle.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central result depends on an unverified realizability assumption: the planned ZLIP ZMP trajectory is only indirectly enforced through a Bézier CoM reference (Eq. 8), so even perfect low-level tracking need not reproduce the planned ZMP dynamics.","rationale":"The reader's weakest assumption identifies exactly the fragility I find most load-bearing: the reduced-order ZLIP plan must be realizable by the full-order closed-loop system. My analysis sharpens this concern by pointing to the specific mechanism—the Bézier CoM reference in Eq. (7)-(8) imposes only endpoint constraints, not the intermediate ZLIP dynamics—so even ideal low-level tracking of the Bézier curve does not imply the planned ZMP trajectory is realized. This is not a contradiction inside the paper, but it is an unproven and unvalidated link between the reduced-order stability analysis and the full-order simulation results. The paper's internal ablations support the qualitative benefit of combining the three strategies, and the deterministic simulations are consistent with the central claim, so I do not see grounds to reject. However, the lack of a tracking guarantee or of even a force-based CoP check means the central claim is not yet fully secured; conditional acceptance with a request for this validation is the appropriate verdict. Since the reader already reached CONDITIONAL with high confidence, my read does not change the verdict.","tokens_in":12324,"tokens_out":8247,"duration_ms":88240,"concrete_test":"Re-run the 130 N sagittal flat-footed push scenario in Section VI-A while logging, at the 1 kHz control rate, the planned p_zmp trajectory from ZLIP-MPC and the actual center of pressure (CoP) computed from MuJoCo contact wrenches. Compute the maximum deviation |p_zmp_actual - p_zmp_planned| and check whether the actual CoP leaves the support polygon at any time. Then replace the Bézier CoM reference in Section V-B with a reference obtained by direct numerical integration of Eq. (4) using the planned (p_zmp_dot, Δ_zmp, T_i) sequences, and rerun the same push. If the full-order response—stabilization success and CoM trajectory—changes materially, the Bézier reconstruction is the weak link and the stability claim requires an explicit tracking guarantee.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing claim is that simultaneously using foot placement, step time, and ZMP/ankle control enlarges the push-recovery envelope (Section I, abstract). The reduced-order planner (ZLIP-MPC) optimizes states (p, L, p_zmp) with controls including p_zmp_dot and instantaneous ZMP jumps. However, the full-order controller never commands ZMP directly: it constructs a desired CoM trajectory as a Bézier polynomial p^d_com(s_i) in Eq. (7), with coefficients constrained in Eq. (8) only to match the current CoM position, the final CoM position, and the final pre-impact L*. This does not constrain the intermediate CoM accelerations to follow the planned ZLIP dynamics, so the actual CoP/ZMP realized by the robot during the FA and OA domains can differ from the planned p_zmp. The closed-loop hybrid system is therefore not guaranteed to inherit the reduced-order stability properties; the paper gives no formal tracking bound and validates the method only with single deterministic MuJoCo runs (Figs. 7-10), without reporting measured CoP/ZMP or hardware results. This is load-bearing because if the realized ZMP leaves the support polygon or the planned L evolution is not tracked, the conclusion that 'only the proposed method can stabilize walking under such extreme disturbances' may be an artifact of the specific Bézier reconstruction and simulator rather than a property of the unified controller.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified push-recovery framework for bipedal walking that combines foot placement, step-time adjustment, and ankle/ZMP control in a single model predictive control (MPC) layer. The reduced-order model, called ZLIP, augments the linear inverted pendulum with a ZMP state and control input, and its step-to-step linear dynamics are used to formulate a small nonlinear program with ZMP support-polygon constraints and step-time constraints. The MPC determines foot placement, step timing, and ZMP trajectories for flat-footed and multi-domain (heel-to-toe) walking. A task-space quadratic program tracks Bézier-parameterized output trajectories on a simulated Cassie robot. Validation is performed in MuJoCo: an ablation study for a 130 N sagittal push and a 300 N lateral push in flat-footed walking, and a 100 N push for multi-domain walking. The paper claims, to the authors' knowledge, that this is the first method to simultaneously employ foot placement, step time adjustment, and ankle torque in an optimization-based framework, and that the proposed method significantly enlarges the push-recovery envelope.","tokens_in":12671,"tokens_out":5441,"duration_ms":55818,"significance":"If the underlying realizability assumptions are met, this paper would be a valuable contribution: it integrates three well-studied recovery strategies into one MPC formulation, exploits closed-form step-to-step dynamics to keep the optimization small, explicitly handles multi-domain contact phases, and demonstrates real-time feasibility (8 ms average solve time at 50 Hz). The ablation study comparing the full method against controllers with one strategy disabled is a useful and honest way to examine the contribution of each mechanism. The main significance is therefore conditional on closing the gap between the reduced-order ZLIP plan and the actual full-order robot behavior, since the paper does not demonstrate that the planned ZMP commands are actually realized.","major_comments":[{"comment":"The realizability of the planned ZLIP ZMP trajectory is not established. Equation (8) constrains only the current CoM position, the pre-impact CoM position, and the pre-impact angular momentum; it does not constrain intermediate CoM accelerations to follow the ZLIP dynamics of Eq. (4). Consequently, even with perfect low-level tracking of the Bézier CoM reference, the realized CoP/ZMP need not match the planned p_zmp, and the closed-loop hybrid system may not inherit the reduced-order stability properties. No measured CoP/ZMP or CoM tracking error is reported in Section VI. I ask the authors to report planned versus realized ZMP and CoM tracking errors in the simulations, and to provide either a formal tracking bound or additional constraints on Eq. (8) that enforce the ZLIP dynamics along the Bézier reference.","section":"V-B, Eq. (8)"},{"comment":"The central claim that \"only the proposed method can stabilize walking under such extreme disturbances\" is supported by one deterministic simulation per condition. There is no systematic characterization of the push-recovery envelope, for example through sweeps over force magnitude, duration, and direction, and no multiple trials with perturbed initial conditions. Since the abstract and Section I claim a significantly enlarged push-recovery envelope, I request a quantitative envelope comparison across controllers with success/failure regions rather than single representative runs.","section":"VI, Figs. 7-10"},{"comment":"The MPC initial condition Xnow omits the current ZMP state and instead assumes that the ZMP follows the nominal reference trajectory from prior work. Because the actual ZMP is not measured on Cassie and the planned ZMP is only indirectly executed through the CoM reference, the reduced-order state used for feedback may diverge from the true robot state under large perturbations. A sensitivity analysis, or the use of an estimator for p_zmp, is needed to support the robustness conclusions.","section":"IV-A"},{"comment":"The ablated controllers are not pure. The \"no foot placement\" condition still allows a 5 cm relaxation around the nominal foot placement, and the \"no ZMP\" and \"no step time\" conditions are implemented by fixing bounds rather than removing the corresponding decision variables. This makes the attribution of the improvement to each individual control input less clean than stated. Please state exactly which decision variables are frozen and verify that the relaxed variants indeed exclude the mechanism under test.","section":"VI-A, Fig. 7"}],"minor_comments":[{"comment":"There is a typo in Section III: \"multi-domian\" should be \"multi-domain\".","section":"III"},{"comment":"The symbol TSS is used in Section VI-B without being defined; it appears to denote the total single-support time (TFA + TUA) and should be defined at first use.","section":"VI-B"},{"comment":"The caption of Fig. 10 lists panels (a), (b), (c), and (e), but the text refers to panel (d); the caption should include panel (d) or the in-text reference should be corrected.","section":"Fig. 10"},{"comment":"The sentence \"A preview of n = 2 steps is used for [35] the ZLIP-based MPC planner\" contains an unexplained citation bracket in the middle of the sentence and should be reworded.","section":"VI-A"},{"comment":"The claim that this is \"the first method that simultaneously employs all three control strategies\" should be qualified relative to existing DCM/Capture-Point step-timing controllers [20]-[23] and constrained-foothold MPC footstep planners [17], which already combine subsets of these mechanisms; please explain the distinguishing features explicitly.","section":"I"},{"comment":"The notation in Eq. (6) would benefit from a brief definition of T^i_j as the newly received duration for domain i, since the superscript/subscript convention is otherwise introduced only implicitly.","section":"V-A, Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The novelty claim should be checked carefully against the DCM-based step-timing literature and MPC footstep planners with ZMP constraints; the distinguishing feature (simultaneous ankle, step-time, and foot-placement optimization in one NLP) must be stated precisely. The paper would also benefit from acknowledging that the ZMP execution is indirect and from providing the requested tracking data, as this is the main load-bearing risk in the reviewer assessment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a real engineering contribution, not a paradigm shift. The ZLIP-MPC cleanly unifies foot placement, step timing, and ZMP/ankle control in a single step-to-step MPC, and the internal ablation shows all three inputs contribute to push recovery in simulation. The main weakness is a gap between the reduced-order plan and full-order execution that is only validated by a few single-run MuJoCo simulations.\n\nWhat's new: the ZLIP model augments the usual LIP state with a ZMP integrator, giving closed-form step-to-step dynamics with ZMP rate and step duration as controls. That allows the MPC to modulate ankle torque, step timing, and foot placement simultaneously. The multi-domain heel-to-toe application is a natural but nontrivial extension. The math is standard linear-system analysis and looks correct. The ablation study (no-ZMP, no-step-time, no-foot-placement) is a decent internal control and supports the claim that all three strategies together beat any pair.\n\nWhere it's soft: first, the realizability concern. The MPC plans a ZMP trajectory, but the low-level controller only tracks a CoM Bézier polynomial constrained at endpoints and final angular momentum (Eq. 8). There is no guarantee the intermediate CoM acceleration matches the planned ZLIP dynamics, so the realized ZMP can differ from the plan. The paper acknowledges Cassie lacks foot force sensors but gives no tracking bound and doesn't report measured ZMP. This is the load-bearing gap behind stability claims. Second, the simulation results are single deterministic runs. No seeds, no error bars, no parameter sweeps. For a robustness paper, that's thin. Third, the \"first\" claim isn't benchmarked against external DCM-based controllers that already combine step placement and timing (e.g., Griffin et al., Khadiv et al.). The internal ablation beats its own baselines but doesn't establish state-of-the-art.\n\nThe citation pattern is fine; reliance on [24] for nominal orbits is a dependency but disclosed. The model derivations are reproducible and the code framework is standard for the community.\n\nBottom line: this deserves a serious referee. It's a well-executed simulation study with a clean formulation and a plausible central mechanism. The realizability gap and lack of statistical rigor are addressable and don't look fatal. If I were an editor, I'd send it to review with a request for stronger validation: multiple seeds, reported ZMP, an external comparison, and ideally hardware or a more formal tracking analysis.","headline":"A clean reduced-order MPC for push recovery that unifies foot placement, step timing, and ZMP control; simulation results are plausible but the full-order tracking gap and single-run validation need attention.","tokens_in":13187,"tokens_out":2640,"would_cite":true,"duration_ms":24921,"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":"The paper claims that robust push recovery during walking requires optimizing foot placement, step timing, and ankle torque together, and gives a real-time MPC controller that does so.","keywords":["push recovery","bipedal locomotion","model predictive control","reduced-order model","zero moment point","multi-domain walking","foot placement","step timing"],"falsifier":"Conduct a hardware push-recovery trial on the same biped under a 130 N sagittal push, measuring the actual ZMP with instrumented feet; if the closed-loop system becomes unstable while the commanded ZMP stays inside the support polygon and the MPC remains feasible, the tracking assumption behind the claim would be refuted.","tokens_in":12104,"feed_emoji":"🤖","tokens_out":10586,"duration_ms":90957,"temperature":0.7,"pith_summary":"In the tested cases, a walking robot recovers from large unknown pushes only when it can use three control levers at the same time: where the next foot lands, when it lands, and how the ankle shifts the center of pressure inside the stance foot. The authors build a reduced-order model they call the ZLIP model, which places the zero moment point as a state inside the standard linear inverted pendulum, and wrap it in a model predictive controller that optimizes foot placement, step duration, and ZMP motion over a short preview horizon. In simulation on a 3D underactuated biped, the full three-lever controller stabilizes pushes that defeat controllers using any two of the three levers, for both flat-footed and multi-domain heel-to-toe gaits. If the claim holds, it unifies previously separate push-recovery strategies into one real-time optimization.","feed_headline":"Three push-recovery levers in one controller beat any pair","feed_subtitle":"A single model predictive controller unifies foot placement, step timing, and ankle torque to enlarge the stable push-recovery envelope.","key_machinery":"The load-bearing object is the ZLIP model, an augmented linear inverted pendulum whose state is $\\xi=[p, L, p_{\\mathrm{zmp}}]^T$, where $p$ is the horizontal center-of-mass position, $L$ is the mass-normalized centroidal angular momentum, and $p_{\\mathrm{zmp}}$ is the horizontal zero-moment-point position relative to the stance pivot. Its continuous dynamics are linear in the ZMP velocity $\\dot{p}_{\\mathrm{zmp}}$, and its discrete transition maps encode foot placement as the swing-foot touchdown position and allow instantaneous ZMP jumps; domain durations are also control inputs. This single model covers flat-footed walking as the special case of a trivial underactuated phase with no ZMP travel, and heel-to-toe multi-domain walking with fully actuated, underactuated, and overactuated phases. The model turns the full hybrid dynamics into a step-to-step nonlinear program whose solution generates foot placement, step timing, and ZMP commands, which are then converted into center-of-mass references and tracked by a task-space quadratic-programming controller.","core_discovery":"The central claim, stated as the paper's key contribution, is that this is the first optimization-based push-recovery framework to simultaneously employ foot placement, step time adjustment, and ankle torque. The ZLIP model represents the center of mass, its mass-normalized angular momentum, and the ZMP position as a linear three-state system, making the ZMP velocity, instantaneous ZMP jumps, and domain durations control inputs. Because the model's dynamics are linear, the MPC can be formulated step-to-step as a small nonlinear program rather than a full transcription, and the ablation studies in the results section show that disabling any one of the three control levers leads to falling under the tested disturbances, while the complete controller recovers from a 130 N sagittal push lasting 0.5 s, a 300 N lateral push, and a 100 N push during multi-domain heel-to-toe walking.","pith_inferences":["The same ZLIP formulation could be extended to full-body humanoids by adding centroidal angular momentum regulation as an additional continuous control input, a direction the paper itself names as future work.","Because the step-to-step NLP avoids discretizing the continuous dynamics, adding preview steps may be relatively cheap; a testable extension is to measure how solve time grows with horizon length and with the number of contact phases.","The reported failure modes suggest a hierarchy in which foot placement is the indispensable recovery lever, while step time and ZMP control add margin when foot placement alone would exceed kinematic limits; the paper does not prove this ordering, so it remains an inference from the ablations.","Since the reduced model treats domain transitions as purely time-based, coupling the ZLIP planner to contact-event or slip feedback is a natural next test for robustness on less ideal surfaces."],"forward_implications":["Removing any one of the three levers--ankle or ZMP control, step-time control, or foot-placement control--makes the robot fall under the tested pushes, while the complete controller recovers.","The MPC shortens the fully actuated and double-support phases in response to a push, a step-time reaction similar to human perturbation responses.","The same framework applies to both flat-footed walking and multi-domain heel-to-toe walking, so the robustness gains are not tied to a single contact mode.","The planner runs at 50 Hz with an average solve time of 8 ms on a desktop CPU with no special optimization, so the approach is feasible for real-time use.","A two-step preview horizon is sufficient for the tested disturbances, and the framework can be extended to more preview steps if needed."],"supporting_citations":[{"why":"Supplies the prior multi-domain walking controller, nominal orbits, and output-construction framework that this work extends.","marker":"[24]"},{"why":"Defines the mass-normalized centroidal angular momentum that appears as a ZLIP state.","marker":"[28]"},{"why":"Provides the optimization framework used to formulate and solve the nonlinear MPC program.","marker":"[29]"},{"why":"Provides the interior-point solver used for the MPC optimization.","marker":"[30]"},{"why":"Describes the 18-DOF underactuated biped platform that the simulations model.","marker":"[31]"},{"why":"Supplies the task-space quadratic-programming controller used for low-level tracking.","marker":"[32]"},{"why":"Provides the open-source simulation environment used for the biped validation experiments.","marker":"[33]"},{"why":"Provides the physics engine that supplies the high-fidelity simulation dynamics.","marker":"[34]"},{"why":"Justifies the two-step preview horizon used in the experiments.","marker":"[35]"}],"fun_headline_variants":["Three levers beat two: MPC unifies push recovery","Bipedal push recovery: foot, timing, ankle in one MPC","First MPC to combine all three push-recovery levers","Disable any lever, robot falls; all three, it recovers","One controller, three levers: robust push recovery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the low-level tracking controller can make the full robot reproduce the reduced-model plan closely enough: the ZLIP commands are realized through a center-of-mass reference, and the paper validates this inheritance of stability only in simulation, with no formal proof and no hardware experiment.","fun_headline_variants_meta":{"raw":{"variants":["Three levers beat two: MPC unifies push recovery","Bipedal push recovery: foot, timing, ankle in one MPC","First MPC to combine all three push-recovery levers","Disable any lever, robot falls; all three, it recovers","One controller, three levers: robust push recovery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000366,"raw_usage":{"total_tokens":1940,"prompt_tokens":889,"completion_tokens":1051,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":965}},"tokens_in":505,"tokens_out":1051,"duration_ms":8653,"temperature":1.0,"reasoning_tokens":965,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:12:16.862082+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Conduct a hardware push-recovery trial on the same biped under a 130 N sagittal push, measuring the actual ZMP with instrumented feet; if the closed-loop system becomes unstable while the commanded ZMP stays inside the support polygon and the MPC remains feasible, the tracking assumption behind the claim would be refuted.","supporting_citations":[{"cited_title":"Multi-domain walking with reduced- order models of locomotion,","cited_arxiv_id":null,"evidence_quote":"Supplies the prior multi-domain walking controller, nominal orbits, and output-construction framework that this work extends."},{"cited_title":"Centroidal dynamics of a humanoid robot,","cited_arxiv_id":null,"evidence_quote":"Defines the mass-normalized centroidal angular momentum that appears as a ZLIP state."},{"cited_title":"https://www.agilityrobotics.com/robots#cassie","cited_arxiv_id":null,"evidence_quote":"Describes the 18-DOF underactuated biped platform that the simulations model."},{"cited_title":"Quadratic Programming for Multirobot and Task-Space Force Control,","cited_arxiv_id":null,"evidence_quote":"Supplies the task-space quadratic-programming controller used for low-level tracking."},{"cited_title":"https://github.com/osudrl/cassie-mujoco-sim, 2018","cited_arxiv_id":null,"evidence_quote":"Provides the open-source simulation environment used for the biped validation experiments."},{"cited_title":"MuJoCo: A physics engine for model-based control,","cited_arxiv_id":null,"evidence_quote":"Provides the physics engine that supplies the high-fidelity simulation dynamics."},{"cited_title":"Two steps is enough: No need to plan far ahead for walking balance,","cited_arxiv_id":null,"evidence_quote":"Justifies the two-step preview horizon used in the experiments."}],"review_version":1}