{"id":"518ddbc1-c09f-468b-b716-fa13aba5104e","arxiv_id":"2509.04722","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-level MPC framework for humanoid walking that optimizes step timing, step length, and ankle torque with ALIP dynamics at the top and a linear arm/torso-extended SRB tracker below, improving push recovery and yaw rejection.","lead":"This paper introduces a hierarchical controller for the Unitree G1 humanoid that plans steps with a nonlinear model and tracks them with a fast linear model, adding arm and torso dynamics. In simulations and hardware tests the controller improved push recovery and reduced yaw sway during walking.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Upper-body yaw-rejection benefit rests on a restrictive arm model and is validated only in simulation; hardware experiments disable upper-body control (Sec. IV-D.3).","rationale":"I considered two candidate concerns: (1) the 36% push-recovery improvement confounds adaptive step timing with a different nominal step period and lacks error bars; (2) the upper-body yaw-rejection benefit relies on a highly simplified arm/torso model and is only demonstrated in simulation. The reader's weakest_assumption was (2). I agree with the reader that (2) is more load-bearing. The upper-body DSRB is a stated contribution (Sec. III-C) and the abstract claims it improves yaw disturbance rejection without qualifying that the supporting experiments are simulation-only. The hardware terrain tests disabled upper-body control (Sec. IV-D.3, Remark 2), so there is no evidence the effect transfers to the physical robot. The model's assumptions—point-mass arms moving only along x, fixed y/z moment arms, no arm joint/torque limits—are precisely the conditions that could break on hardware. If a full-order simulation with realistic arm constraints or a hardware test shows no benefit, the paper's novelty would be substantially reduced. The 36% issue is real but secondary; even if the statistic is inflated, adaptive step timing is qualitatively demonstrated in hardware push recovery. Therefore the conditional verdict should stand, pending hardware or full-order validation of the upper-body controller.","tokens_in":12469,"tokens_out":15756,"duration_ms":154488,"concrete_test":"Run the Sec. IV-B yaw disturbance experiment on the Unitree G1 with the upper-body DSRB control enabled, using the same Mz range (30-150 Nm for 0.05 s) during 0.3 m/s walking, and compare pelvis yaw deviations against the SRB MPC baseline. If DSRB does not yield a significant reduction, the upper-body contribution is unsupported. Alternatively, in MuJoCo add arm joint torque/position limits and full arm inertia (y/z) to the G1 model; if the DSRB advantage over SRB disappears under these limits, the x-only point-mass arm model is too restrictive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's claim that the DSRB mid-level MPC reduces pelvis yaw deviations (Sec. IV-B) is a central novelty, yet it is supported only by simulation. The DSRB model in Sec. III-C assumes each arm is a point mass constrained to move along x with fixed y/z moment arms (F_LA,y=F_LA,z=0), and the torso-lower-body interaction is a single z torque. The yaw disturbance tests in Fig. 4 use this model, but the hardware experiments in Sec. IV-D.3 explicitly disabled upper-body control. If the real G1 arm inertias, joint limits, or torque limits prevent the commanded x-axis arm forces from being realized, the predicted yaw rejection improvement may not transfer to hardware. The abstract states the result without the simulation-only caveat, overclaiming generalization. This is load-bearing because the upper-body DSRB is one of the paper's main contributions and is unvalidated on the actual robot.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical model predictive control framework for humanoid locomotion. A high-level NMPC optimizes step lengths, step periods, and ankle torques using the step-to-step dynamics of the ALIP model at 40 Hz; the resulting plan is converted into references for a 500 Hz linear MPC based on a decomposition of the single rigid body dynamics into lower body, torso, and arms (DSRB). The manuscript reports simulation studies showing that the DSRB model reduces pelvis yaw deviation after yaw moment disturbances and that adaptive step periods increase push-recovery success by 36%, plus hardware push-recovery and terrain-walking experiments on the Unitree G1.","tokens_in":12775,"tokens_out":4674,"duration_ms":48141,"significance":"If validated, the proposed hierarchical architecture is practically important: it demonstrates that nonlinear step-timing and step-length optimization can run on an onboard mini-PC at 40 Hz while a convex MPC runs at 500 Hz, and the DSRB model is an elegant way to incorporate simplified upper-body actuation into a linear MPC. The hardware demonstrations on indoor/outdoor terrain support feasibility and robustness. However, the two headline quantitative claims rest on modest simulation evidence, and the upper-body benefit is not demonstrated on hardware. The paper would be strengthened by more rigorous experimental reporting and by qualifying simulation-only claims.","major_comments":[{"comment":"The 36% improvement in push-recovery success is not supported by the reported data. Each force combination was tested only 5 times, with a binary success/failure outcome; with n=5, a single trial changes the success rate by 20 percentage points, so no statistically meaningful comparison can be made. The figure also appears to conflate two factors: the fixed-step condition uses T=0.35 s while the adaptive condition uses a desired period of 0.4 s with bounds 0.25–0.5 s. The improvement may be due to the lower nominal step frequency rather than to adaptation. Please report per-cell success counts, confidence intervals (or a formal test), and run control experiments that vary the fixed step period and the desired adaptive period independently.","section":"§IV-C, Fig. 5"},{"comment":"The central claim that upper-body control improves yaw disturbance rejection is supported only by simulation. The hardware terrain experiments in Sec. IV-D.3 explicitly state that upper-body control was disabled, so no hardware evidence validates the DSRB model. Moreover, the DSRB model in Sec. III-C assumes each arm is a point mass moving only along the x-axis, with F_LA,y=F_LA,z=0 and fixed moment arms r_LA,y/r_LA,z, and that the torso–lower-body interaction is a single z-axis torque. There is no demonstration that these assumptions are realizable on the G1 hardware, e.g., under joint limits, arm inertia, and torque limits. The abstract and conclusion currently state the benefit without this simulation-only caveat. Please either add hardware yaw-disturbance experiments with upper-body control enabled or, at minimum, substantially rephrase the claims and add a sensitivity/discussion sec","section":"§IV-B and §IV-D.3"},{"comment":"The push-recovery simulation experiments are not described in enough detail to be reproducible or to assess the effect's magnitude. The text reports 'external force (F_x: –600 – 600 N, F_y: 0 – 400 N) was applied for 0.1 s' and 'the robot was walking in place,' but does not state how the force was applied in simulation (e.g., direction relative to the walking frame, the exact timing distribution, or the robot's state distribution at push). The success rate depends strongly on push phase; with only 5 trials, the phase coverage is likely poor. Please provide a detailed experimental protocol, including the number of trials per condition, the phase sampling method, and the variance across trials.","section":"§IV-C, general experimental methodology"}],"minor_comments":[{"comment":"The state x_SRB in Eq. (2) has 15 components (p, Θ, v, ω, g), while Eq. (13) uses an identity matrix I_13×13. This is a dimension mismatch that should be corrected or clarified.","section":"§II-D, Eq. (13)"},{"comment":"The figure shows raw data and binned means/standard deviations, but no sample size per bin is given. The binning in 10 Nm steps is coarse; please report the number of trials per bin and consider plotting individual trajectories or medians with interquartile ranges.","section":"§IV-B, Fig. 4"},{"comment":"The notation for the stance foot position p_STF,xy,k and the use of k_n in the following section is not fully defined; please clarify the indexing and the relationship between step index k and MPC node n.","section":"§III-B.1, Eq. (23)"},{"comment":"The phrase 'the upper body control improved the yaw disturbance rejection' should be qualified as a simulation result, since hardware experiments did not exercise upper-body control.","section":"Abstract"},{"comment":"There are several minor typographical issues in equations, e.g., the multiplication dots in Eq. (16) and the layout of Eq. (28). A careful proofread would improve readability.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"I see no fundamental theoretical error in the framework; the concern is that the two headline contributions are empirically under-supported. The paper could become acceptable if the authors strengthen the experimental validation and reframe the simulation-only claims. Given the paper's scope, I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid engineering paper. The genuinely new piece is the ALIP S2S NMPC that optimizes step period, step length, and ankle torque together, plus the \"decomposed\" SRB-MPC that adds arms and torso while keeping the QP linear. I don't see that exact combination in the cited prior work, and the derivation is clear. If you work on humanoid locomotion, this is worth knowing about.\n\nWhat is good: the architecture is sensible—40 Hz NLP for the nonconvex step timing/placement, 500 Hz linear MPC for forces, and a clean IK/torque mapping. The simulation results are consistent with the qualitative claims, and the hardware terrain runs show the overall controller can handle grass, stone, uneven mats, and step-downs on a real G1. The paper also states its modeling assumptions plainly; there is no hidden fitting of constants to make the result come out.\n\nWhere it is soft: the headline 36% improvement is simulation-only, uses only five trials per force pair, and no confidence intervals are reported. Worse, the fixed-step baseline uses T=0.35 s while the adaptive case has desired T=0.4 s, so the comparison is not apples-to-apples. The upper-body yaw-rejection benefit is also simulation-only; the hardware terrain experiments explicitly disable upper-body control. Given that the DSRB model restricts arms to x-axis point masses with fixed y/z moment arms and a single torso-z torque, the transfer to the real G1's arm inertia and joint limits is genuinely unproven. The abstract should at least scope this claim. The stress-test note about this is fair.\n\nOne more minor point: the paper's conclusion admits the high-level ALIP model doesn't capture torso/arm dynamics, which is fine, but it means the \"hierarchical\" coupling is one-way. That is a limitation, not a fatal flaw.\n\nBottom line: the central architecture is coherent, the math checks out as far as I followed it, and the evidence supports the qualitative story even if the headline number is oversold. This deserves peer review. My recommendation: send it to review, but the authors should be asked to fix the baseline comparison, add error bars or more trials, and explicitly scope the upper-body results as simulation-only.","headline":"Solid, useful humanoid locomotion controller; the headline robustness number is real but oversold by simulation-only results and a confounded baseline comparison.","tokens_in":13212,"tokens_out":2448,"would_cite":true,"duration_ms":26146,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a hierarchical ROM-MPC with adaptive step timing and arm/torso-aware mid-level control raises push-recovery success by 36% and cuts pelvis yaw drift under torso twists.","keywords":["humanoid locomotion","model predictive control","reduced-order models","angular momentum linear inverted pendulum (ALIP)","adaptive step timing","push recovery","upper-body control","yaw disturbance rejection"],"falsifier":"Run the robot at 0.3 m/s on hardware, apply a 30–150 Nm yaw moment to the torso for 0.05 s, and compare pelvis yaw recovery with DSRB-MPC versus SRB-MPC; if the yaw deviation curves overlap, the upper-body contribution claim fails. Equally, log commanded versus measured arm force during recovery: if the arm joints cannot deliver the commanded x-axis forces within their torque limits, the mechanism is not realized.","tokens_in":12384,"feed_emoji":"🦿","tokens_out":4675,"duration_ms":44554,"temperature":0.7,"pith_summary":"This paper is trying to establish that a humanoid robot can be made more robust for real-world walking by splitting locomotion control into two predictive layers built on simplified models: a slower nonlinear planner that chooses step lengths, step durations, and ankle torques using the angular-momentum ALIP model, and a fast linear MPC that tracks the resulting plan while also using arm accelerations and torso rotation to resist perturbations. If true, it would mean humanoid locomotion controllers can keep the tractability of reduced-order models while recovering some of the stabilizing behavior usually reserved for whole-body control. The paper reports a 36% increase in push-recovery success rate with adaptive step timing over a fixed step period, and reduced pelvis yaw deviations when the robot is twisted about the vertical axis, validated in simulation and on the Unitree G1 humanoid. The upper-body yaw-rejection result is shown in simulation; the hardware terrain experiments run with upper-body control disabled.","feed_headline":"Adaptive step timing lifts humanoid push recovery by 36 percent","feed_subtitle":"A two-level MPC lets a robot change step length and cadence mid-walk and uses arms and torso to resist yaw twists.","key_machinery":"The load-bearing object is the step-to-step (S2S) ALIP dynamics in (11), which maps pre-impact angular-momentum states across steps under step length, step period, and ankle torque. It carries the high-level NMPC cost and constraints, letting the planner vary step timing and placement. The second mechanism is the Decomposed SRB (DSRB) dynamics in (28), an extension of single-rigid-body dynamics with torso yaw and two arm point masses whose x-axis acceleration produces equal-and-opposite forces and moments; the linearity comes from fixing the arm y/z positions and forces to zero, so the moment from each arm is a constant cross product times a scalar force. This keeps the mid-level MPC a quadr","core_discovery":"The central claim is that an ALIP-based step-to-step nonlinear MPC can solve for foot placement, step timing, and ankle torque together, and that its plan can be converted into references for a linear SRB MPC whose model is decomposed to include arms and torso without losing linearity. Concretely, the paper shows the dynamics in (28): arms are modeled as point masses constrained to move along the x-axis, generating a force and coupled moments, and a torso yaw actuator generates an opposite moment on the lower body. With this Decomposed SRB model, the mid-level MPC can command arm forces and torso torques to reject yaw moments. The authors report that adaptive step periods in the range 0.25–0","pith_inferences":["An implication the authors leave implicit is that the yaw-rejection benefit should transfer best to disturbances whose moment-arm geometry matches the assumed fixed r_LA,y and r_LA,z; a testable extension is to add asymmetric torso loads or off-axis pushes and observe when arm joint torque limits saturate.","The 36% push-recovery gain in simulation may understate or overstate hardware performance depending on push phase; a hardware experiment sweeping push timing uniformly over the step cycle would give a tighter estimate of the mean recovery probability.","Because the high-level ALIP planner ignores arm and torso dynamics, feeding DSRB state limits and upper-body constraints into the high-level planner is a natural next step; one concrete test is whether adding arm-state bounds preserves yaw rejection under larger twists.","The arm yaw-rejection mechanism depends on the controller being able to command x-axis arm forces fast enough; logging commanded versus measured arm force during a recovery would reveal whether hardware joints can actually realize the simulated benefit."],"forward_implications":["A humanoid can absorb pushes by shortening or lengthening steps on the fly rather than keeping a fixed cadence, improving recovery without heavier whole-body optimization.","The DSRB formulation gives a computationally cheap way for upper-body motion to actively cancel yaw moments, reducing reliance on stance-foot friction for torsion disturbances.","The 40 Hz high-level / 500 Hz mid-level split suggests the full stack can run on low-power onboard compute while keeping real-time quadratic-program solves.","The same S2S ALIP plan can feed different mid-level models, making the high-level planner a reusable interface for step placement and timing.","The approach is robust across uneven outdoor and indoor terrains despite the ROM assumptions, pointing toward deployment in varied real environments."],"supporting_citations":[{"why":"Supplies the ALIP model and its angular-momentum-about-contact-point formulation, the foundation for the step-to-step dynamics used in the high-level planner.","marker":"[10]"},{"why":"Provides the HLIP model and period-1/period-2 orbit definitions used to compute desired pre-impact ALIP states from velocity commands.","marker":"[9]"},{"why":"Supplies the force-and-moment-based SRB-MPC formulation whose dynamics are linearized and extended into the mid-level DSRB controller.","marker":"[21]"},{"why":"Establishes the standard convex SRB-MPC approach for legged robots that the paper decomposes to add arm and torso dynamics.","marker":"[22]"},{"why":"Demonstrates a hierarchical MPC that adapts gait frequency for humanoid push recovery, providing the direct baseline and motivation for adaptive step periods.","marker":"[17]"},{"why":"Shows arm motion planning for improved disturbance recovery, supporting the paper's use of arm forces to improve robustness.","marker":"[19]"},{"why":"Provides the swing-foot trajectory blending method used by the low-level joint mapper to smooth transitions during stepping.","marker":"[23]"}],"fun_headline_variants":["Two-level MPC lets humanoids adapt steps and use arms for balance","Adaptive step timing in MPC boosts push recovery by 36%","Hierarchical MPC with arm control improves humanoid locomotion","Robust humanoid walking via reduced-order MPC and arm dynamics","Step, timing, and arm torques: unified MPC for humanoid robustness"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The mid-level DSRB model assumes each arm is a point mass that can only be accelerated along the x-axis with fixed y/z lever arms, and that the whole upper-body effect on the pelvis is a single z-axis torque; if real arm inertia, joint limits, or off-axis forces violate this, the predicted yaw rejection may not appear on hardware.","fun_headline_variants_meta":{"raw":{"variants":["Two-level MPC lets humanoids adapt steps and use arms for balance","Adaptive step timing in MPC boosts push recovery by 36%","Hierarchical MPC with arm control improves humanoid locomotion","Robust humanoid walking via reduced-order MPC and arm dynamics","Step, timing, and arm torques: unified MPC for humanoid robustness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000149,"raw_usage":{"total_tokens":1025,"prompt_tokens":737,"completion_tokens":288,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":199}},"tokens_in":481,"tokens_out":288,"duration_ms":3335,"temperature":1.0,"reasoning_tokens":199,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:56:09.367233+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the robot at 0.3 m/s on hardware, apply a 30–150 Nm yaw moment to the torso for 0.05 s, and compare pelvis yaw recovery with DSRB-MPC versus SRB-MPC; if the yaw deviation curves overlap, the upper-body contribution claim fails. Equally, log commanded versus measured arm force during recovery: if the arm joints cannot deliver the commanded x-axis forces within their torque limits, the mechanism is not realized.","supporting_citations":[{"cited_title":"Force-and-moment-based model predictive control for achieving highly dynamic locomotion on bipedal robots,","cited_arxiv_id":null,"evidence_quote":"Supplies the force-and-moment-based SRB-MPC formulation whose dynamics are linearized and extended into the mid-level DSRB controller."},{"cited_title":"Dynamic locomotion in the mit cheetah 3 through convex model-predictive control,","cited_arxiv_id":null,"evidence_quote":"Establishes the standard convex SRB-MPC approach for legged robots that the paper decomposes to add arm and torso dynamics."},{"cited_title":"Adapting Gait Frequency for Posture-regulating Humanoid Push-recovery via Hierarchical Model Predictive Control","cited_arxiv_id":"2409.14342","evidence_quote":"Demonstrates a hierarchical MPC that adapts gait frequency for humanoid push recovery, providing the direct baseline and motivation for adaptive step periods."},{"cited_title":"Humanoid arm motion planning for im- proved disturbance recovery using model hierarchy predictive control,","cited_arxiv_id":null,"evidence_quote":"Shows arm motion planning for improved disturbance recovery, supporting the paper's use of arm forces to improve robustness."},{"cited_title":"Dynamic walking on highly underactuated point foot humanoids: Closing the loop between hzd and hlip,","cited_arxiv_id":null,"evidence_quote":"Provides the swing-foot trajectory blending method used by the low-level joint mapper to smooth transitions during stepping."}],"review_version":1}