{"id":"398dc9a9-5a08-4f36-957a-db7ebf0e058d","arxiv_id":"2505.23499","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A preview-control-based centroidal controller generates and stabilizes humanoid multi-contact motions at sub-millisecond cost, demonstrated in simulation.","lead":"This paper replaces computationally expensive model predictive control with classic preview control to generate and stabilize whole-body multi-contact trajectories for a humanoid robot, running in about 0.1 to 0.4 milliseconds per control step. In simulation, the method carries an HRP-5P robot through walking, handrail stairs, and ladder climbing without solving a constrained optimization at each step.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The angular-momentum approximation (L≈Iω with constant diagonal I, Euler-angle-rate ω) is validated only for two small-orientation motions; the cartwheel that would stress it is not executed, so the 'general multi-contact motion' claim is not yet supported.","rationale":"The reader's weakest_assumption correctly identifies the angular-momentum approximation as the key vulnerability in the central claim. My independent reading of the full text confirms this: the preview control's angular subsystem uses a constant diagonal inertia and Euler-angle rates, and Section V-F validates it only on two motions with small base inclinations. The reported mean error of 3.1 kg m^2/s between the full centroidal angular momentum and the model is non-negligible, and the reduction to 0.011 relies on an angular momentum task that is not part of the proposed controller. The cartwheel, which would stress the approximation, is explicitly not executed. The computational speed claim, in contrast, is well supported by the timing measurements in Table IV and the simple matrix-multiplication structure. The parameter-contradiction issue noted by the reader is real but secondary; it affects a contribution claim, not the core feasibility of the method. Since the demonstrated motions all have small orientation changes, the concern does not invalidate those specific results, but it does limit the generality claim. The reader's CONDITIONAL verdict appropriately requires additional validation of the angular model; my stress-test does not change that verdict, so I recommend UNCHANGED.","tokens_in":13005,"tokens_out":6468,"duration_ms":64934,"concrete_test":"Execute the Section V-D cartwheel centroidal trajectory on the simulated HRP-5P using a whole-body IK that includes an orientation task (and optionally the angular momentum task described in Section V-F) so that the planned base orientation can actually be tracked. If the robot falls or the centroidal tracking error diverges during the large-rotation phase, the angular approximation is invalid for motions with large orientation changes, directly undermining the 'general multi-contact motion' claim. If execution is infeasible, compute ||I_all(q) ˙q − Iω|| along the simulated trajectories of all five Fig. 4 motions; any motion with a peak error comparable to the commanded angular momentum scale (e.g., > 3.1 kg m^2/s) indicates the model is inaccurate for that motion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that preview control with approximated centroidal dynamics can stably generate a variety of multi-contact motions. The angular subsystem (Eqs. 6–7) assumes L = Iω with a constant diagonal inertia, ignores joint momentum and pose dependence of the centroidal momentum matrix, and drops the Euler-angle kinematic map K_Euler. Section V-F validates this approximation only for bipedal walking and handrail stairs, reporting a mean ||I_all ˙q − Iω|| of 3.1 kg m^2/s even in those benign cases; the error only drops to 0.011 after an angular momentum task is added to the whole-body IK, which is not part of the preview controller itself. No error analysis is provided for the other demonstrated motions (Fig. 4 C/D/E) that involve large hand forces, moving contacts, or different joint configurations. The one motion explicitly designed to stress the approximation, the cartwheel in Section V-D with a full base-link revolution, is not executed on the robot. The claim of handling 'general multi-contact motion' therefore rests on an angular model whose accuracy is unverified precisely in the regime where it is likely to fail. If the angular model is wrong, the planned wrench and the post-hoc wrench projection are systematically biased, and the stabilizer must compensate, which may explain why only near-upright motions succeed. This is load-bearing because the method deliberately omits the constraints that MPC enforces, so the model must be accurate for the wrench projection to be meaningful.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a centroidal trajectory generation and stabilization framework for humanoid multi-contact motion. Instead of solving a constrained MPC problem at each step, the method uses linear preview control on the 6-DoF centroidal dynamics, approximating the centroidal angular momentum as L ≈ Iω with a constant diagonal inertia matrix and using Euler-angle rates as the angular velocity. The planned resultant wrench is projected onto the contact constraint manifold via a QP, and a centroidal PD feedback law together with per-limb damping control provides stabilization. The authors validate the approach in simulation on a HRP-5P model for bipedal walking and four multi-contact scenarios, report sub-millisecond computation times for a 2-s preview horizon, and show a planned cartwheel trajectory as a stress test for the angular model. The paper's central claim is that this combination achieves fast, general multi-contact motion control without the computational burden of constraint-enforcing MPC.","tokens_in":13394,"tokens_out":3528,"duration_ms":37522,"significance":"If the central claims are confirmed, the paper makes a useful contribution: it offers a computationally inexpensive alternative to MPC for centroidal multi-contact control, with a clear algorithmic structure and concrete simulation demonstrations. The derivation of the connection between the proposed feedback law and DCM-based bipedal control in the appendix is a nice theoretical bridge that could help practitioners tune gains. The paper also provides detailed parameter tables and simulation protocols, which makes the method reproducible in principle. However, the significance is tempered by gaps in the validation of the angular-momentum approximation, which is load-bearing for the claimed generality of the method, and by the fact that some controller parameters and reference trajectories are motion-specific, contradicting the stated parameter-free claim. These issues are addressable, but they need to be fixed before the broader claims can be accepted.","major_comments":[{"comment":"The validation of the angular-momentum approximation (Eq. 6) is insufficient to support the claim of general multi-contact motion. The paper reports a mean error ||I_all qdot − Iω|| of 3.1 kg m^2/s even for bipedal walking and handrail stairs — two motions with relatively small base-link orientation changes — and the error is reduced to 0.011 only after adding an angular-momentum task to the whole-body IK, which is not part of the proposed controller. No error metrics are reported for the other demonstrated motions in Fig. 4(C), (D), and (E), which involve large hand forces, moving contacts, or different joint configurations, and the cartwheel stress test in Section V-D is planned but not executed. Because the preview controller and the wrench projection rely on this approximation, an unquantified model bias could be absorbed by the feedback stabilizer, making the success of the near-upright motions a weak witness for the model's validity. Please provide per-motion error statistics for all demonstrated motions and either execute the cartwheel or give a principled argument, with numerical evidence, for why the approximation holds in the large-orientation regime.","section":"Section V-F and Section III-A4"},{"comment":"The claim that the method handles bipedal walking and multi-contact motions 'without changing the parameters' is contradicted by the manuscript's own tables. Table II changes the hand damping gain Kd to 1000 for the wall-walking motion and to 50000 for the ladder motion, and Table III replaces the centroidal feedback gains KP and KD entirely for the ladder motion. In addition, the reference CoM trajectories are motion-specific: the stair motion uses a 50-mm inward offset and the ladder motion uses a 0.4-m offset with a 0.1-m forward adjustment, chosen per motion. Parameter tuning per motion is not inherently a flaw, but the claimed parameter invariance is a key selling point. Please either revise the claim to reflect the disclosed tuning, or demonstrate a single fixed parameter set that works across all presented motions, documenting how the reference offsets are generated from a common rule.","section":"Section V-C, Tables II and III, and Contribution (ii)"},{"comment":"The comparison with constrained MPC is not sufficient to support the concluding statement that 'there is no significant difference between the two methods.' The comparison reports mean projection errors of the planned and desired wrenches, but these are not accompanied by statistical measures, tracking-error metrics, or stability-margin indicators. Moreover, the MPC baseline from reference [5] may include different feedback or posture tasks, making it unclear what exactly is being compared. Please either present a more quantitative comparison (e.g., CoM tracking RMSE, contact-force tracking error, or a robustness perturbation test) or soften the conclusion to a statement about comparable projection-error magnitudes in the specific tested scenarios.","section":"Section V-G"}],"minor_comments":[{"comment":"The heading 'F . V alidation of Rotational Motion Approximation' contains stray spaces; it should be 'F. Validation of Rotational Motion Approximation'.","section":"Section V-F heading"},{"comment":"The legend of Figure 8(B) lists 'planned force' twice; the second occurrence should likely be 'actual force' or another quantity, and the figure would be clearer if the legend entries were corrected.","section":"Figure 8(B)"},{"comment":"The phrase 'the reference CoM trajectory is determined by a simple rule; the horizontal position is the center of the supporting foot...' is clear, but the sentence in Section V-C that says 'the reference CoM trajectories are represented by piecewise-constant functions' could be confusing because the actual reference trajectories shown in Figures 5-7 are step-like but not piecewise constant in the vertical direction; please clarify.","section":"Section V-B"},{"comment":"In the reporting of projection errors, it is not always clear which quantity the pair '(2.2 N and 4.4 Nm)' refers to — force and moment parts of the same vector norm or separate norms; please define the notation explicitly.","section":"Section V-G"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound in its core construction, but the gap between the claimed generality and the provided validation is substantial. The angular-momentum approximation issue is the most serious: the paper itself flags the strong assumptions and then validates them only on the two most benign motions, with the cartwheel left unexecuted. The parameter-tuning contradiction is also a credibility issue for the 'without changing parameters' contribution. Both are fixable within the scope of a revision. I would be willing to look at a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis is a real engineering contribution, not a repackaging of Kajita 2003. Murooka et al. redefine the preview-control output as CoM plus resultant wrench instead of ZMP, extend the state to 6-DoF base orientation with an approximate angular model, and combine it with wrench projection and centroidal PD feedback. The result is a controller that updates in 0.1–0.4 ms over a 2 s / 400-sample horizon, and they show stable simulation of five multi-contact motions, from handrail stairs to moving environments. That speed claim is credible and useful.\n\nThe paper is honest about its own limitations: the angular approximation L ≈ Iω with constant diagonal inertia is stated as an assumption, and Section V-F validates it only for walking and handrail stairs. The reported mean error of 3.1 kg m²/s for ‖I_all q̇ − Iω‖ is not negligible, even in those benign cases. The cartwheel motion, which would actually stress the approximation, is generated but not executed. So the \"general multi-contact motion\" claim outruns the evidence: the paper demonstrates several multi-contact motions, which is already a good result.\n\nThe tuning story is slightly oversold too. The text says the same parameters are used everywhere, but Tables II and III list motion-specific hand Kd and ladder feedback gains, plus per-motion CoM offsets. They disclose this openly, so it is a wording problem rather than a hidden failure. The MPC comparison is limited to two motions and the projection error is sometimes similar to MPC, which is fair, but the MPC baseline is their own prior work and no code or data are released for independent reproduction.\n\nThe math itself is standard preview control with a clean DCM consistency check in the appendix. Nothing circular. The wrench projection is a least-squares fit. I believe the central method works as described for the class of near-upright motions demonstrated.\n\nWho is this for? Anyone building fast centroidal controllers for humanoid multi-contact locomotion. It deserves peer review, but the authors should be asked to either execute a large-orientation motion, report approximation error for all demonstrated motions, or temper the \"general\" claim. For a revision, I would also ask for the cartwheel or an equivalent stress test as a planned trajectory at minimum.\n\nMy recommendation: engage with this, send it to review, and push on the angular validation. It is a solid submission with one soft spot that is addressable.\n\nBest","headline":"Solid engineering: fast preview-control centroidal multi-contact stabilization that mostly delivers, but the angular momentum approximation is validated on only benign motions and the 'general multi-contact' claim outruns the evidence.","tokens_in":13902,"tokens_out":2570,"would_cite":true,"duration_ms":25557,"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 preview control, not MPC, can generate and stabilize humanoid multi-contact motion at 0.1–0.4 ms per control cycle with a 2-second lookahead.","keywords":["humanoid robots","multi-contact motion","centroidal dynamics","preview control","stabilization control","wrench distribution","receding horizon","whole-body control"],"falsifier":"Run the controller on a motion with large joint swings and a fast base rotation—for example, a cartwheel or backflip—and compute, online or offline, the time series of $\\|I_{\\mathrm{all}}\\dot{q} - I\\omega\\|$ (and $\\|I\\omega - I\\dot{\\alpha}\\|$); if the error grows substantially beyond the 3.1 kg m²/s mean reported for the validated walking and stair motions, or if the simulated robot falls during execution, the central claim that this approximation supports general multi-contact motion is falsified.","tokens_in":12822,"feed_emoji":"🤖","tokens_out":7959,"duration_ms":75171,"temperature":0.7,"pith_summary":"This paper claims that humanoid multi-contact motion—bipedal walking, climbing handrail stairs, climbing a vertical ladder, walking with hands on a wall, and balancing on moving floors and walls—can be planned and stabilized online from centroidal dynamics using preview control instead of full model predictive control. In simulation, the whole control update, including feedback stabilization and contact-wrench distribution, takes 0.1–0.4 ms while looking 400 samples (2 s) ahead, roughly one to two orders of magnitude faster than the MPC alternatives it compares against. That speed matters because the centroidal trajectory can be regenerated every control tick, which is what lets the robot absorb disturbances and follow contacts that move. The paper claims that the contact constraints omitted by preview control can be handled afterward by projecting the planned wrench onto the contact constraint manifold, and the simulations support that claim for a variety of motions.","feed_headline":"Preview control steers multi-contact humanoid motion in under 1 ms","feed_subtitle":"One controller with the same gains walks, climbs, and balances on moving contacts in simulation.","key_machinery":"The load-bearing mechanism is the preview-control law for a triple-integrator plant, applied per axis to both the linear and angular centroidal dynamics. With state $x = (c, \\dot c, \\ddot c)$ and input $u = \\dddot c$, the optimal input is $$u^*[k] = -K_{\\mathrm{fb}} x[k] + \\sum_{i=1}^{N_h} K_{\\mathrm{ff}}[i] y_{\\mathrm{ref}}[k+i],$$ where the gains depend only on the constant robot mass (and the approximated constant diagonal inertia for the angular part), so the runtime cost is one matrix multiplication regardless of the 400-sample horizon. The angular channel rests on the approximation $L \\approx I\\omega$ with constant diagonal $I$ and Euler-angle rates in place of true angular velocity. Around this core, a QP projects the planned wrench onto the contact wrench cone $w_i = G_i\\lambda_i$, $\\lambda_i \\ge 0$, and damping control at each limb end tracks the distributed contact wrenches.","core_discovery":"The central claim, stated on the paper's own terms, is that a 6-degree-of-freedom centroidal trajectory—center of mass plus base orientation—can be generated online by solving, for each of the six axes, an independent preview-control problem, with the CoM and the resultant wrench (force and moment combined with gravity) as outputs instead of the ZMP used in classical bipedal preview control. For each axis the dynamics are a triple integrator: state (position, velocity, acceleration), input jerk, output (position, resultant wrench); the optimal input is a fixed state feedback plus a weighted sum of future reference outputs, so at runtime it is a single matrix multiplication. The planned wrench is then projected onto the contact constraint manifold, and a centroidal PD feedback law plus per-limb damping control turns it into stable whole-body motion. The paper further shows that the proposed feedback law contains DCM-based bipedal balance control as a special case, which gives a principled way to choose its gains.","pith_inferences":["Beyond the paper, the preview-plus-projection recipe may transfer to other systems whose dynamics reduce to centroidal form, such as quadrupeds or manipulators in contact, wherever a rough reference and a constraint-projection layer are available.","The cartwheel results validate only trajectory generation, not execution; if whole-body tracking were added, the constant-inertia and Euler-angle-rate approximations would be the first place to look for failure, since the paper itself leaves execution out of scope.","The reference force is always set to zero in the paper, so an immediate testable extension is to feed nonzero reference forces (for example, when the robot must push or carry a load) and measure whether tracking degrades.","One could also replace the angular approximation with a better online inertia estimate and keep the same preview structure; the speed advantage would likely survive because only the gains would need updating."],"forward_implications":["A single set of controller gains, except for a few hand-damping values, covers bipedal walking and the tested multi-contact motions, which means the method scales to new contact schedules without re-tuning the whole pipeline.","Because the gains are fixed as long as mass (and inertia) stay constant, the online computation does not grow with the preview horizon; looking 2 s ahead costs the same as looking 0.5 s ahead.","The feedback law's equivalence to DCM control for bipedal walking gives a closed-form way to convert a desired DCM gain into the centroidal feedback gains, transferring tuning intuition from bipedal walking to multi-contact motion.","Compared with constrained MPC on the same tasks, the planned and feedback-modified wrenches show projection errors of the same scale, so for motions that include feedback the two approaches are not significantly different despite the huge cost difference."],"supporting_citations":[{"why":"Supplies the preview-control optimal-input formula (5) that reduces runtime to a single matrix multiplication.","marker":"[11]"},{"why":"Provides the bipedal ZMP preview-control formulation whose triple-integrator state equation the paper extends to 6-DoF multi-contact with CoM-plus-wrench outputs.","marker":"[21]"},{"why":"Defines the centroidal dynamics and the centroidal momentum matrix $I_{\\mathrm{all}}$ that the angular approximation (6) simplifies.","marker":"[1]"},{"why":"Serves as the constrained-MPC multi-contact baseline for the computation-time and trajectory-quality comparisons.","marker":"[5]"},{"why":"Gives the contact wrench parametrization $w_i = G_i \\lambda_i$ used in both wrench projection and distribution.","marker":"[19]"},{"why":"Provides the limb-end damping control law (15) that tracks the distributed contact wrenches.","marker":"[23]"},{"why":"Defines DCM-based bipedal balance control, whose proportional form the proposed centroidal feedback generalizes and whose gain correspondence is derived in the appendix.","marker":"[16]"}],"fun_headline_variants":["Centroidal preview control speeds multi-contact humanoid motion","Preview control cuts compute cost for multi-contact humanoid tasks","One preview loop stabilizes walking, climbing, and balancing","Preview control replaces MPC for real-time multi-contact humanoid control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the robot's angular momentum can be modeled as a constant diagonal inertia times the base-link angular velocity, with joint motion and orientation-rate corrections ignored; if a motion involves large limb swings or fast orientation changes, the planned angular trajectory and wrench projection are built on a model that may be far off, and the robot can fall.","fun_headline_variants_meta":{"raw":{"variants":["Centroidal preview control speeds multi-contact humanoid motion","Preview control cuts compute cost for multi-contact humanoid tasks","One preview loop stabilizes walking, climbing, and balancing","Preview control replaces MPC for real-time multi-contact humanoid control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000524,"raw_usage":{"total_tokens":2473,"prompt_tokens":828,"completion_tokens":1645,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":444,"completion_tokens_details":{"reasoning_tokens":1577}},"tokens_in":444,"tokens_out":1645,"duration_ms":17046,"temperature":1.0,"reasoning_tokens":1577,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:44:38.626607+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the controller on a motion with large joint swings and a fast base rotation—for example, a cartwheel or backflip—and compute, online or offline, the time series of $\\|I_{\\mathrm{all}}\\dot{q} - I\\omega\\|$ (and $\\|I\\omega - I\\dot{\\alpha}\\|$); if the error grows substantially beyond the 3.1 kg m²/s mean reported for the validated walking and stair motions, or if the simulated robot falls during execution, the central claim that this approximation supports general multi-contact motion is falsified.","supporting_citations":[{"cited_title":"Design of an optimal controller for a discrete-time system subject to previewable demand,","cited_arxiv_id":null,"evidence_quote":"Supplies the preview-control optimal-input formula (5) that reduces runtime to a single matrix multiplication."},{"cited_title":"Biped walking pattern generation by using preview control of zero-moment point,","cited_arxiv_id":null,"evidence_quote":"Provides the bipedal ZMP preview-control formulation whose triple-integrator state equation the paper extends to 6-DoF multi-contact with CoM-plus-wrench outputs."},{"cited_title":"Centroidal dynamics of a humanoid robot,","cited_arxiv_id":null,"evidence_quote":"Defines the centroidal dynamics and the centroidal momentum matrix $I_{\\mathrm{all}}$ that the angular approximation (6) simplifies."},{"cited_title":"Model preview control in multi-contact motion-application to a humanoid robot,","cited_arxiv_id":null,"evidence_quote":"Serves as the constrained-MPC multi-contact baseline for the computation-time and trajectory-quality comparisons."},{"cited_title":"Multi-contact motion planning and control,","cited_arxiv_id":null,"evidence_quote":"Gives the contact wrench parametrization $w_i = G_i \\lambda_i$ used in both wrench projection and distribution."},{"cited_title":"Biped walking stabilization based on linear inverted pendulum tracking,","cited_arxiv_id":null,"evidence_quote":"Provides the limb-end damping control law (15) that tracks the distributed contact wrenches."}],"review_version":1}