{"id":"6e28c6d8-1fda-4674-aa22-5f6516c8b2b3","arxiv_id":"2505.11495","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A simulated humanoid uses its arms to brace against walls and recover from pushes up to 100 N while walking, surviving 70.6% of tested push scenarios versus 16.8% for a stepping-only controller.","lead":"This paper describes a controller that lets a simulated humanoid robot keep walking after being shoved by bracing its arms against a wall. A generalist reader might care because surviving unexpected pushes is a key step toward safe humanoid robots in human-centered spaces.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 70.6% safe-set gain is never shown to come from arm/wall bracing; without an arm-disabled ablation, the central mechanism is unverified.","rationale":"The reader's weakest_assumption was the SRB reduction's massless-limb premise. That is related and relevant, but the more decision-relevant gap is that the evaluation never isolates the arm-bracing mechanism that the abstract names as the 'key innovation'. If an arm-disabled ablation leaves performance roughly unchanged, the paper's contribution statement is overstated even though the combined controller works. My proposed test is feasible in MuJoCo and would settle the attribution. I am not claiming the results are fabricated; rather, the causal claim is underdetermined by the current experiment. The paper also has no code release, which makes independent replication harder, but that by itself is not a correctness objection. Since the concern is about missing evidence rather than demonstrated error, it keeps the reader's CONDITIONAL verdict rather than moving it to reject or accept.","tokens_in":11176,"tokens_out":7469,"duration_ms":81353,"concrete_test":"Re-run the full 1,440-push benchmark with the identical SRB-MPC-HLIP controller and all gains/thresholds fixed, but force the hand-contact force u_h to zero and never add the wall-contact constraints in Eq. (17), so the arms cannot brace. If the safe-set fraction remains close to 70.6%, arm/wall bracing is not the source of the claimed improvement; if it falls to near the 16.8% HLIP-alone baseline, the bracing mechanism is confirmed. Report per force/height/speed cells, not just the aggregate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that arms bracing against walls are what enable the improved push recovery, but the only comparison reported (Section IV-C) is full SRB-MPC-HLIP versus 'pure HLIP'. That comparison cannot separate the contribution of wall-contact forces from the other components that change simultaneously: the SRB-MPC's foot-force optimization, the push-detection threshold (Section III-A: 0.4 m/s and 0.2 rad/s, 'based on emperical testing'), and the step-frequency increase from 3 Hz to 5 Hz. If these components, rather than the hand forces, are responsible for the 16.8% to 70.6% improvement, the paper's stated key innovation is not supported. The lack of per-cell success counts also leaves open whether the 'up to 100 N / 0.5 m/s' phrase describes a single configuration or a robust region. The paper itself flags the SRB model's omission of limb-induced disturbances only as future work (Section V), so the full-order fidelity of the planned hand-contact forces for the hardest pushes is also unquantified. A single ablation disabling the hand-contact force entirely, with all other modules and gains fixed, would directly test the causal mechanism at the center of the paper.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a unified push-recovery and locomotion controller for a 24-DOF, 20 kg humanoid, combining a Single Rigid Body model predictive controller (SRB-MPC) with Hybrid Linear Inverted Pendulum (HLIP) stepping control. The robot uses its arms to brace against walls detected via LiDAR when a push is detected, and the MPC re-plans hand and foot contact forces while HLIP adjusts step timing and placement. The authors evaluate the approach in MuJoCo over 1,440 push scenarios (forces 30-100 N, heights 0.1-0.5 m, commanded speeds 0-0.5 m/s, two push sides, three push times), reporting that the SRB-MPC-HLIP controller survives 70.6% of scenarios versus 16.8% for HLIP alone, plus a qualitative angled-wall and multi-directional push demonstration. The paper also includes tracking comparisons and a no-wall push test while stepping in place.","tokens_in":11425,"tokens_out":4957,"duration_ms":46962,"significance":"If the results are supported, the paper would provide a useful modular architecture: a convex MPC formulation with multi-contact wall bracing, a principled HLIP-based footstep adaptation, and an unusually systematic perturbation sweep. Strengths include the explicit derivation of the SRB-MPC dynamics and constraints, the 1,440-case evaluation protocol, the use of a high-fidelity simulator, and the fact that no parameter was fitted to the reported safe-set metric. However, the central causal claim about arm/wall bracing needs a direct ablation, and several controller parameters and per-cell results are not reported; these gaps currently prevent the quantitative claims from being fully assessed.","major_comments":[{"comment":"The paper's headline contribution is that the arms brace against walls, but the only quantitative comparison is the full SRB-MPC-HLIP controller versus pure HLIP. That comparison cannot separate the contribution of the hand-contact force from the simultaneous changes in foot-force optimization, push-detection thresholds, step-frequency increase (3 Hz to 5 Hz), and wall detection. Please add an ablation that disables the hand-contact force (uh = 0 in Eq. (5)) while keeping every other module and tuning fixed, and report the same 1,440-case safe-set metric. Without this, the improvement cannot be attributed to wall bracing.","section":"Section IV-C compared with Sections III-A and III-B"},{"comment":"The phrase 'recover from pushes up to 100 N for 0.2 s while walking at commanded speeds up to 0.5 m/s' is ambiguous because the aggregate 70.6% safe set (Fig. 4) does not show whether the extreme cells (100 N at 0.5 m/s, including each height, side, and phase combination) individually succeed. Please provide per-condition success counts or a heatmap over the swept grid, and state whether the extreme values describe a single successful trajectory or a robust region.","section":"Section IV-C"},{"comment":"The push-detection thresholds (0.4 m/s and 0.2 rad/s) are described as based on 'emperical testing,' and the MPC cost weights Q and R, the HLIP gain K, and the contact-force bounds Ffmax, Fhmin, Fhmax are not reported. Because these parameters determine when recovery is activated and how forces are allocated, the 70.6% result is not reproducible and its sensitivity to these tunings is unknown. Please report the numeric values or release the code/configuration, and include a sensitivity study for the detection thresholds.","section":"Sections III-A and III-D"},{"comment":"Eq. (6) is derived under the assumption that the robot's mass and inertia are concentrated in the torso and that limb dynamics are negligible, and the paper itself lists 'internal disturbances caused by the movements of the limbs' as future work. For the largest pushes in the sweep, the discrepancy between the SRB-predicted contact forces and the full-order MuJoCo response is unquantified. Please compare planned versus actual hand and foot forces in simulation, or otherwise bound the model error, so that the wall-bracing forces are credible at the extremes of the tested range. This is a correctness-risk concern, not a claim of circularity.","section":"Sections II-B and V"}],"minor_comments":[{"comment":"There are typos: 'emperical' should be 'empirical' in Section III-A, and 'feed-foward' should be 'feed-forward' in Section III-G.","section":"Sections III-A and III-G"},{"comment":"The reported increase is numerically inconsistent: 16.8% to 70.6% is a factor of 4.20, which is a 320% increase, not a 420% increase; the caption of Fig. 4(a) says a '225% larger safeset' while the forces shown are 20 N and 30 N. Please harmonize these numbers.","section":"Figure 4 and Section IV-C"},{"comment":"Please define the friction coefficient μ and the default MuJoCo friction settings, since the friction cones in Eq. (17) depend on them.","section":"Section IV-A"},{"comment":"The HLIP feedback gain K in Eq. (23) is said to stabilize A - BK, but the dimension of the HLIP state and the construction of K are not specified; a brief definition would improve reproducibility.","section":"Section III-F"},{"comment":"The paper does not report error bars or repeated stochastic trials; if the simulator has any nondeterminism, please state the number of seeds used for each condition.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The paper is within the scope of a robotics venue and the core architecture is plausible. My main concern is that the central mechanism, arm bracing against walls, is not isolated by an ablation; a modest set of ablations and parameter disclosures would likely make it publishable. I have no concerns about novelty disclosure or citation behavior."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the punchline: this is a credible engineering integration with a real result—a simulated 20 kg humanoid survives 70.6% of a 1,440-push battery while walking at commanded speeds up to 0.5 m/s, versus 16.8% with pure HLIP. But the paper's central claim, that the arms bracing against walls are what enable the improvement, is not actually demonstrated. The comparison is whole-controller versus whole-controller, and several other things change at once: the SRB-MPC foot-force optimization, the push detection threshold, and the stepping frequency (3 to 5 Hz). Without an arm-disabled ablation, the 420% safeset gain can't be attributed to the stated key innovation. This is the main soft spot, and it's a genuine one.\n\nWhat is new: combining SRB-MPC with HLIP for reactive arm-wall contact during locomotion. Prior work either used arms without locomotion (Khazoom and Kim) or offline sampling (Marcucci et al.), so this integration is new. The MPC derivation in Section III is coherent, and the evaluation protocol is systematic: force magnitude, height, side, walking speed, and gait phase swept in a grid. The tracking comparison in Fig. 3 is also useful.\n\nWhat it does well: the authors are honest about the SRB's limitations—they list internal limb disturbances as future work—and the push detection thresholds are described as empirical rather than presented as principled. The angled-wall and multi-directional push demos are a nice addition.\n\nSoft spots beyond the missing ablation: no code or parameter values, so the reported numbers are hard to build on; no repeated stochastic trials, so 70.6% is a point estimate from single MuJoCo runs; and the abstract's 'up to 100 N / 0.5 m/s' reads as a robust region when it is actually the boundary of the aggregate safeset. These are fixable in a revision. The SRB model's massless-limb assumption is a real limitation but it applies equally to both controllers, so it does not undercut the comparison.\n\nWho is this for: people working on humanoid push recovery or using environment contacts for locomotion. It deserves a serious referee, but the reviewers should ask for an arm-disabled ablation, a sensitivity analysis on the detection thresholds and cost weights, and either code or a full parameter table. If those come back clean, it becomes a solid within-subfield contribution.","headline":"The integration is real and the evaluation is systematic, but the paper never isolates the arm-wall bracing from the other controller changes, so its central claim is unverified.","tokens_in":11981,"tokens_out":3278,"would_cite":false,"duration_ms":30201,"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":"A reduced-order controller lets a walking humanoid brace against walls and recover from torso pushes up to 100 N.","keywords":["humanoid push recovery","reduced order models","model predictive control","Hybrid Linear Inverted Pendulum","Single Rigid Body model","multi-contact locomotion","wall bracing","simulation"],"falsifier":"Run the same 1,440-push battery on hardware (or in a full-order simulation that retains arm and leg inertia) with the wall within reach. If the fraction of survived pushes falls to near the 16.8% HLIP-only level, or if measured hand contact forces deviate systematically from the MPC-commanded values, the central claim that SRB-based bracing causes the recovery improvement would be refuted.","tokens_in":10950,"feed_emoji":"🤖","tokens_out":4995,"duration_ms":45760,"temperature":0.7,"pith_summary":"This paper argues that a humanoid robot can keep walking through strong pushes by treating nearby walls as extra points of support, using its arms to brace while the legs adjust their steps. The authors combine two established reduced-order models—the Single Rigid Body (SRB) model, which plans contact forces on a torso with lumped mass, and the Hybrid Linear Inverted Pendulum (HLIP) model, which sets step lengths and timing—into one model-predictive controller that switches into a recovery mode when a push is detected. In simulation of a 24-degree-of-freedom, 20 kg humanoid, the combined controller survives 70.6% of 1,440 push scenarios (up to 100 N for 0.2 s at walking speeds up to 0.5 m/s), while the HLIP-only baseline survives 16.8%. The paper takes this as evidence that multi-contact, wall-assisted recovery can be planned in real time from reduced-order models rather than requiring whole-body optimization or learning.","feed_headline":"Humanoid braces against walls to survive 100 N pushes","feed_subtitle":"Combining two reduced-order models lifts push-recovery success from 17% to 71% in simulation.","key_machinery":"The load-bearing object is the SRB-MPC-HLIP controller, a model-predictive controller built on two reduced-order models. The SRB model (Equation (6) and its discretized form in Equation (11)) treats the robot as a single rigid body with mass and inertia concentrated at the torso, acted on by contact forces and moments at the hands, left foot, and right foot; the HLIP model (Equations (22)–(23)) supplies step-to-step dynamics that prescribe swing-foot placement. The MPC layer solves the quadratic program of Equation (18) with the contact constraints of Equation (17), including hand-on-wall friction cones, and maps optimal contact wrenches to joint torques through Jacobians. The mechanism that carries the argument is the joint activation of hand contact (via inverse kinematics to the wall) and step-pattern adjustment (higher step frequency, wall-aware step limits) whenever the predicted CoM states deviate beyond set thresholds.","core_discovery":"The central claim is that combining SRB-MPC with HLIP dynamics lets a walking humanoid use its arms to brace against walls and still recover from pushes that would defeat a stepping-only controller. The SRB-MPC layer solves, at 250 Hz or faster, a quadratic program for hand and foot contact forces that satisfies friction-cone and force-bound constraints; the HLIP layer raises step frequency from 3 Hz to 5 Hz during recovery and lengthens steps in the push direction to enlarge the support polygon. The two layers are coupled through a push detector that compares predicted center-of-mass velocity and angular velocity against reference values (0.4 m/s and 0.2 rad/s thresholds) and through forward-kinematics knowledge of wall distance. On the 1,440-case push battery, the combined controller lifts the safe-set fraction from 16.8% to 70.6%, a 420% increase, and it also keeps torso angular velocities closer to zero during nominally commanded walking. The authors further report recovery from multi-directional pushes against walls slanted inward at 5 degrees.","pith_inferences":["If the same architecture transfers to hardware, walls, railings, and furniture could be treated as a single unified contact resource during locomotion, which would change how humanoid deployment sites are prepared.","A natural next experiment would be pushes from behind or pushes that occur when no wall is within arm reach, since the current recovery mode is gated on arm workspace reaching a wall; the paper does not report those cases.","One could test the SRB assumption directly by adding mass to the arms or legs and rerunning the push battery; if the safe-set fraction drops toward the HLIP baseline, the torso-lumped-mass idealization is the binding constraint.","The 0.4 m/s and 0.2 rad/s detection thresholds are tuned empirically; a learned or adaptive threshold could trade false alarms against missed pushes, an extension the paper lists as future work."],"forward_implications":["A 20 kg humanoid with 24 degrees of freedom can survive a 100 N, 0.2 s torso push while walking at 0.5 m/s, provided a reachable wall exists.","The same controller, without retraining, handles walls slanted inward by 5 degrees and pushes arriving from either side.","Worst-case omnidirectional pushes without any wall are rejected up to 30 N, against 20 N for HLIP alone, so some robustness gain comes from stepping even before arm contact.","Recovery is fast enough for real-time control: the MPC QP runs at 250 Hz and the step period shortens from 3 Hz to 5 Hz during recovery.","The safe-set increase from 16.8% to 70.6% indicates that environmental bracing, not just stepping, is what closes most of the gap."],"supporting_citations":[{"why":"Establishes the stepping-based push-recovery baseline that the combined controller extends with arm contacts.","marker":"[15]"},{"why":"Supplies the SRB-MPC convex formulation (Equations (6) and (11)) used for contact-force planning.","marker":"[22]"},{"why":"Gives the step-to-step dynamics approximation that motivates using HLIP for robust stepping.","marker":"[8]"},{"why":"Provides the HLIP gait synthesis and stepping stabilization used to generate swing foot trajectories.","marker":"[9]"},{"why":"Supplies the high-fidelity physics simulator used for all reported experiments.","marker":"[23]"},{"why":"Solves the quadratic program that implements the MPC layer.","marker":"[31]"},{"why":"Provides the line-fitting method used to detect wall distance from LiDAR scans.","marker":"[29]"},{"why":"Supplies the swing-foot smoothing technique used to avoid excessive oscillations.","marker":"[32]"}],"fun_headline_variants":["Wall-bracing arms help humanoids recover from 100N pushes","Humanoid uses walls to recover from pushes, 4x better success","Reduced-order models boost humanoid push recovery to 71%","Bracing against walls: humanoid push recovery up 420%","Wall-bracing boosts humanoid push recovery success to 71%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The plan assumes the robot's mass and inertia are concentrated in the torso, so arms and legs act as massless force generators; if limb dynamics contribute meaningfully to balance during large pushes, the planned contact forces will not match the real robot's behavior.","fun_headline_variants_meta":{"raw":{"variants":["Wall-bracing arms help humanoids recover from 100N pushes","Humanoid uses walls to recover from pushes, 4x better success","Reduced-order models boost humanoid push recovery to 71%","Bracing against walls: humanoid push recovery up 420%","Wall-bracing boosts humanoid push recovery success to 71%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000721,"raw_usage":{"total_tokens":3230,"prompt_tokens":932,"completion_tokens":2298,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2206}},"tokens_in":548,"tokens_out":2298,"duration_ms":13759,"temperature":1.0,"reasoning_tokens":2206,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:51:34.253148+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 1,440-push battery on hardware (or in a full-order simulation that retains arm and leg inertia) with the wall within reach. If the fraction of survived pushes falls to near the 16.8% HLIP-only level, or if measured hand contact forces deviate systematically from the MPC-commanded values, the central claim that SRB-based bracing causes the recovery improvement would be refuted.","supporting_citations":[{"cited_title":"Push recovery by stepping for humanoid robots with force controlled joints,","cited_arxiv_id":null,"evidence_quote":"Establishes the stepping-based push-recovery baseline that the combined controller extends with arm contacts."},{"cited_title":"Robust disturbance rejection for robotic bipedal walking: System-level-synthesis with step-to-step dynamics approximation,","cited_arxiv_id":null,"evidence_quote":"Gives the step-to-step dynamics approximation that motivates using HLIP for robust stepping."},{"cited_title":"qpoases: A parametric active-set algorithm for quadratic program- ming,","cited_arxiv_id":null,"evidence_quote":"Solves the quadratic program that implements the MPC layer."},{"cited_title":"Weighted line fitting algorithms for mobile robot map building and efficient data representation,","cited_arxiv_id":null,"evidence_quote":"Provides the line-fitting method used to detect wall distance from LiDAR scans."},{"cited_title":"Dynamic walking on highly underactuated point foot humanoids: Closing the loop between hzd and hlip,","cited_arxiv_id":null,"evidence_quote":"Supplies the swing-foot smoothing technique used to avoid excessive oscillations."}],"review_version":1}