{"id":"b558a4dc-43b4-4fde-9ca9-21e0c8efdd8a","arxiv_id":"2501.12594","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A three-stage optimization framework using a reaction mass pendulum, QP mapping, and whole-body dynamics generates 1.0 m forward jumps and a 0.5 m high jump for a humanoid robot.","lead":"This paper presents a three-stage optimization pipeline that plans a humanoid robot's forward jump, first on a simplified pendulum model, then in joint space, then on a whole-body model. The authors report forward jumps of about 1.0 meter and a 0.5 meter high platform jump on an adult-sized humanoid, with trajectories generated in under 10 seconds.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Whole-body stage's planar single-leg symmetry reduction is unverified; without bilateral GRF/CoP data, dynamic consistency of the claimed 1.0 m jump is not established.","rationale":"The reader's weakest assumption—the planar single-legged symmetry reduction—is indeed the most load-bearing concern. The framework's mathematical pipeline is coherent, and the reported experiments and simulations are genuine evidence that the robot can perform forward jumps. However, the physical validity of the optimized trajectory depends on exact bilateral symmetry and sagittal-only dynamics during contact-rich phases. The paper provides no bilateral force data, no repeated-trial statistics, and no full-model comparison, while simultaneously acknowledging model-plant mismatch, visible CoM/CAM errors, and an undescribed bias added to the optimization goal. These admissions, located in Secs. VI-B and VI-C, limit how much the experiments can support the central claim. Because the concern is not a logical contradiction but an unverified precondition, the appropriate outcome remains CONDITIONAL acceptance pending the proposed force-sensor validation. I therefore leave the reader's verdict unchanged.","tokens_in":15907,"tokens_out":4867,"duration_ms":59170,"concrete_test":"Using the robot's existing six-axis force/torque sensors (Sec. VI-A), record left- and right-foot GRFs and CoPs for at least ten forward jumps generated by the same optimized trajectory, during both launch and landing phases. If the peak left-right vertical-force asymmetry exceeds 10% of total vertical force, or if either foot's CoP leaves its own convex hull while the total CoP remains inside the support polygon, the planar symmetric assumption is violated and the whole-body stage must be rerun with a full 20-DOF model to assess the resulting trajectory differences.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the 3-step framework generates executable forward jumps of roughly 1.0 m distance and 0.5 m height—rests on the whole-body stage (Sec. V) optimizing a planar 5-link single-legged model whose leg mass and inertia are the sum of the two legs (Sec. II-C) and then mirroring that trajectory to both legs. This reduction assumes exact left-right symmetry, no roll or yaw motion, and equal ground reactions at both feet. The launch and flight phases are executed open-loop in position mode (Sec. VI-B), so unmodeled asymmetry is neither compensated nor observable in the reported data. The paper itself reports 'non-ideal contact' vibrations (Sec. VI-B), visible CoM/CAM tracking errors, and that the robot 'does not reach the specified CoM height' (Sec. VI-C); it then attributes the discrepancy to measurement error without supporting evidence. The added 'bias to the optimization's goal' is also undescribed. If bilateral ground reactions differ during the contact-rich launch or landing, the planned CoP constraint (Eq. 22) and the CAM trajectory are not dynamically consistent, and the successful trials cannot be cleanly credited to the framework. This is load-bearing because this reduction is the only bridge from the optimized model to the physical robot.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-stage offline trajectory optimization framework for generating forward jump motions on an adult-sized humanoid robot. Stage 1 uses a 2D SRMP model to optimize centroidal momentum, inertia, and CoP trajectories for the launching and flight phases; Stage 2 maps these trajectories into joint space via QP solvers on a 20-DOF full-body model; Stage 3 performs a whole-body optimization on a planar 5-link model whose leg mass and inertia are aggregated from the two legs, with the resulting trajectory mirrored to both legs. The optimized trajectory is executed open-loop in position mode during launching and flight, with a landing controller inherited from prior work. The central claims are that the three-stage framework generates agile forward jumps of about 1.0 m distance and 0.5 m height, that the whole optimization takes less than 10 seconds, and that inertia shaping during flight achieves favorable landing posture.","tokens_in":16118,"tokens_out":4020,"duration_ms":43115,"significance":"If fully validated, the framework would be a practical and fast trajectory-optimization pipeline for humanoid jumping, with a sensible hierarchical decomposition: SRMP-based momentum/inertia generation, QP-based joint-space mapping, and whole-body refinement. The physical robot experiments in Sec. VI-C show real forward jumps, which is nontrivial evidence that the optimized trajectories are executable, and the paper usefully identifies inertia shaping during flight as a key ingredient for landing posture. However, the validation is incomplete: there are no error bars, only selected trials are shown, the flight-phase CoM/CAM are not directly measured, bilateral ground-reaction data are absent, and admitted discrepancies between reference and actual performance are rationalized rather than quantified. The central methodological bridge from the planar symmetric whole-body model to the physical two-legged robot is therefore not yet demonstrated.","major_comments":[{"comment":"The whole-body optimization uses a planar 5-link single-legged model whose leg mass and inertia are the sum of the two legs, and the resulting trajectory is mirrored to both legs. This assumes exact left-right symmetry and identical ground reactions during launch and landing. The paper never reports bilateral force/torque sensor data or CoP tracks from the experiments, so the dynamic consistency of the planned CoP constraint (Eq. 22) and the CAM trajectory with the real two-legged motion is not established. Please provide per-foot ground-reaction force and CoP measurements for the reported trials, or explicitly state the symmetry assumption and justify it with data.","section":"Sec. II-C and Sec. V"},{"comment":"The paper states that the robot does not reach the specified CoM height and that the actual CAM deviates from the reference, then suggests these discrepancies 'might be attributed to measurement errors' without supporting evidence. Because the flight-phase CoM and CAM are not measured (no external motion capture or state estimator), the claimed 1.0 m distance and 0.5 m height cannot be directly verified from the presented data. Please report multiple trials with error bars, measured takeoff velocity, measured flight-phase trajectories, or clearly limit the claim to selected successful trials.","section":"Sec. VI-C, Fig. 14"},{"comment":"The simulation section says that model mismatch can be handled by 'adding bias to the optimization's goal,' but no definition, numerical value, or tuning procedure for this bias is given. This is a load-bearing detail for reproducibility, since the bias appears to be the mechanism by which the framework compensates for dynamic-model differences. Specify where the bias enters the optimization (objective or constraint), how it was selected, and whether it was tuned on the experimental trials.","section":"Sec. VI-B"},{"comment":"The paper claims that the whole optimization process takes less than 10 seconds, but no timing measurements, solver specifications, problem sizes, or hardware details are reported anywhere in the results. Since this is one of the stated contributions and is essential for the claimed online deployability, please provide quantitative timing data for each of the three optimization stages.","section":"Sec. I and Sec. VI"},{"comment":"The simulation validation shows visible differences between reference and actual CoM position/velocity and CAM curves, attributed to dynamic model differences, but no quantitative error metric or acceptance threshold is given. Given that the simulation also uses modified dynamic parameters and open-loop execution, please quantify the tracking errors and show that the achieved jump distance and height match the targets within a defined tolerance.","section":"Sec. VI-B, Fig. 12"}],"minor_comments":[{"comment":"There are several typos and grammatical errors, including 'researches' and 'constracted' in Sec. I, 'the relationship the relationship' in Sec. IV-A, and 'cased' in Sec. VI-B. The paper would benefit from a careful language edit.","section":"Throughout"},{"comment":"The subplot captions in Fig. 12 are not self-contained; please specify in the caption which physical quantities are plotted and clarify the meaning of 'Ref' and 'Act'.","section":"Fig. 12"},{"comment":"The symbol φ is used for the barbell radius in the SRMP configuration (Eq. 1) and also appears in derived quantities such as Eq. 4; the notation should be checked for consistency to avoid confusion between radius and other angular variables.","section":"Sec. II-A and Sec. III-A"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' prior vertical-jump work for the landing controller, the simulator platform, and several equations in Secs. III-V. This is not disqualifying, but the novelty relative to that work should be more crisply stated. A more serious concern is the gap between the claimed validation and the presented evidence: the central symmetry reduction is unverified, and key flight-phase quantities are not measured. If the authors can add bilateral contact data, repeated trials with statistics, and a clear description of the bias term, the paper could become publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a real engineering result, not a breakthrough. A humanoid robot actually jumps forward about a meter and onto a 0.5 m platform using a three-stage trajectory optimizer that couples momentum, inertia, and posture. That is worth knowing. The combination of SRMP-based launch optimization, a centroidal-dynamics QP mapping, and whole-body refinement is new for this platform and motion class, and the sub-10-second solve time is plausible given the staged decomposition. The paper is strongest when it shows the optimized inertia shaping during flight: the robot retracts its legs and arms to swing the feet to a good landing point, which is exactly the kind of behavior that momentum alone cannot produce.\n\nThe soft spots are mostly in validation. The whole-body stage assumes planar bilateral symmetry by modeling one leg with the summed mass/inertia of two, then mirroring the trajectory. The paper does not report per-foot ground reaction forces or any measure of asymmetry, so we cannot see whether the planned CoP and CAM are dynamically consistent during launch or landing. The robot succeeded, which suggests the assumption is not fatal, but it is unverified. The model-plant mismatch is acknowledged and then waved away with a statement that a bias can be added to the optimization goal; that bias is not described or tested. The runtime claim is unsupported by profiler data. And there are no error bars or repeated-trial statistics; only three successful experiments are shown.\n\nI do not think these issues sink the paper. The central claim — that the framework can generate executable forward jumps — is supported by real hardware. The symmetry concern is real but not clearly load-bearing, because the robot did jump the distance. The missing statistics and the vague bias term are more about completeness than correctness. For a hardware paper, this is a normal level of evidence, though a reviewer should push for more.\n\nWho should read this? People working on humanoid jumping, dynamic maneuvers, or trajectory optimization with simplified models. It will be more useful as an engineering reference than as a source of general theory. It deserves peer review; I would accept it with the expectation of a major-revision request that addresses the validation gaps.","headline":"A real robot doing 1m forward jumps is the bottom line; the paper is a credible engineering contribution with some under-specified validation.","tokens_in":16701,"tokens_out":3173,"would_cite":true,"duration_ms":34178,"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 a 3-step trajectory optimization framework, using an SRMP model, joint-space QP mapping, and whole-body optimization, generates executable forward jumps of 1.0 m distance and 0.5 m height on a humanoid robot, with…","keywords":["humanoid robot","forward jump","trajectory optimization","reaction mass pendulum","inertia shaping","centroidal angular momentum","center of pressure","whole-body optimization"],"falsifier":"Use external motion capture to measure the robot's actual takeoff velocity and flight-phase pitch rotation during a forward jump: if the horizontal center-of-mass velocity at takeoff differs from the planned value by more than the landing controller can absorb, or if the robot's roll or yaw grows beyond a small threshold, the planar symmetry assumption behind the whole-body stage is violated and the claimed 1.0 m jump is not reproducible.","tokens_in":15649,"feed_emoji":"🤖","tokens_out":4562,"duration_ms":45030,"temperature":0.7,"pith_summary":"This paper tries to show that a humanoid robot's forward jump can be planned by splitting the problem into three staged optimizations, each using a different dynamic model, so that the whole plan is computed in seconds rather than minutes. The first stage treats the robot as a pendulum whose inertia can change, jointly shaping momentum and body posture. The second maps that plan into joint motions. The third refines it on a whole-body planar model, and the result sends a 42 kg robot 1.0 m forward and 0.5 m high, in simulation and hardware. If right, this makes agile jumping a practical offline-planning routine rather than a heavy whole-body optimization.","feed_headline":"3-step optimizer plans a humanoid's 1.0 m forward jump in 10 seconds","feed_subtitle":"Staged SRMP-to-whole-body pipeline coordinates posture, momentum and inertia shaping for a 42 kg humanoid.","key_machinery":"The central object is the SRMP (static reaction mass pendulum) model, a 2D pendulum whose center of mass is split into a barbell of two masses separated by a controllable radius $\\varphi$, giving total rotational inertia $\\rho = 2 m_p \\varphi^2$. This scalar encodes the robot's body inertia, letting the optimizer shape inertia during flight while total angular momentum is conserved. The center-of-pressure constraint (Eq. 22) couples linear and angular momentum to the support region, and the flight-phase optimization treats the spatial inertia $\\rho^R_{\\text{total}}$ as the control variable to adjust the landing posture.","core_discovery":"The central claim is that the coupling between body posture and centroidal angular momentum during launch, and the inertia shaping that controls rotation during flight, can be optimized together by a split-level approach. A static reaction mass pendulum (SRMP)—a pendulum whose endpoint mass is a barbell of adjustable radius—captures both the momentum and the rotational inertia of the robot, and its optimization produces reference trajectories for momentum, inertia, and center of pressure. The paper asserts that these references, after a joint-space mapping step and a final whole-body optimization on a planar 5-link model, yield a trajectory that the real robot can execute open-loop through launch and flight, landing with the feet placed for a stable touchdown.","pith_inferences":["The planar symmetry reduction suggests an obvious stress test: deliberately perturb the robot's initial posture in roll and yaw and see if the open-loop trajectory still lands stably; the paper does not report such a test.","The 10-second runtime depends on the QP mapping providing a good guess; this staged warm-start strategy could be reused in other whole-body optimal control problems, such as running takeoffs or stair hopping.","The paper's own admission that model error can be pre-compensated by biasing optimization goals hints that a sensitivity analysis—how much target bias per unit of model error—would turn the framework into a robust tuning tool.","If the inertia-shaping flight phase is the key new element, one could test it in isolation by executing only the flight-phase trajectory during a no-jump, in-place rotation maneuver."],"forward_implications":["Total optimization time under 10 seconds makes the framework feasible for online re-planning when jump targets change.","Inertia shaping during flight gives the robot control over landing orientation without breaking angular momentum conservation, so landing foot placement can be planned.","The center-of-pressure constraint links linear and angular momentum to the support region, so the optimized launch trajectory is executable without foot tipping.","The same split-level approach can be applied to other sagittal-plane jumps, such as backflips, by changing the target rotation."],"supporting_citations":[{"why":"Introduces the reaction mass pendulum (RMP) model that the SRMP is derived from, supplying the barbell representation of body inertia.","marker":"[32]"},{"why":"Defines centroidal dynamics and the spatial inertia split that the flight-phase optimization relies on.","marker":"[35]"},{"why":"Provides the relationship between center of pressure and the derivatives of linear and angular momentum used in the CoP constraint.","marker":"[34]"},{"why":"Previous vertical jump work that supplies the landing-phase controller and the centroidal dynamics equations reused in the joint-space mapping.","marker":"[2]"},{"why":"Demonstrates momentum-aware optimization including inertia for quadrupedal jumping, the template for coordinating momentum and inertia in this framework.","marker":"[31]"},{"why":"Defines composite rigid body inertia, which the paper invokes to justify treating inertia shaping as the only posture-control channel during flight.","marker":"[21]"}],"fun_headline_variants":["3-step jump planner: humanoid leaps 1.0 m forward in seconds","Optimized 3-step launch gives humanoid a meter-long jump","SRMP+QP pipeline: humanoid jumps 1.0 m in three steps","Humanoid's 1 m jump: 3-step optimization, 42 kg, open-loop","1 m jump via 3-step SRMP-QP-wholebody optimization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The final optimization runs on a planar 5-link model that merges the two legs into one, so the whole plan assumes the robot's left and right legs move identically and that roll and yaw dynamics are negligible; if the real robot breaks that symmetry, the planned center-of-pressure and landing postures are not dynamically consistent.","fun_headline_variants_meta":{"raw":{"variants":["3-step jump planner: humanoid leaps 1.0 m forward in seconds","Optimized 3-step launch gives humanoid a meter-long jump","SRMP+QP pipeline: humanoid jumps 1.0 m in three steps","Humanoid's 1 m jump: 3-step optimization, 42 kg, open-loop","1 m jump via 3-step SRMP-QP-wholebody optimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000585,"raw_usage":{"total_tokens":2714,"prompt_tokens":876,"completion_tokens":1838,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":1732}},"tokens_in":492,"tokens_out":1838,"duration_ms":11868,"temperature":1.0,"reasoning_tokens":1732,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:00:33.785723+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use external motion capture to measure the robot's actual takeoff velocity and flight-phase pitch rotation during a forward jump: if the horizontal center-of-mass velocity at takeoff differs from the planned value by more than the landing controller can absorb, or if the robot's roll or yaw grows beyond a small threshold, the planar symmetry assumption behind the whole-body stage is violated and the claimed 1.0 m jump is not reproducible.","supporting_citations":[{"cited_title":"Reaction mass pendulum (RMP): An explicit model for centroidal angular momentum of humanoid robots,","cited_arxiv_id":null,"evidence_quote":"Introduces the reaction mass pendulum (RMP) model that the SRMP is derived from, supplying the barbell representation of body inertia."},{"cited_title":"Motion having a flight phase: Experiments in- volving a one-legged robot,","cited_arxiv_id":null,"evidence_quote":"Provides the relationship between center of pressure and the derivatives of linear and angular momentum used in the CoP constraint."},{"cited_title":"Vertical jump of a humanoid robot with cop-guided angular momentum control and impact absorption,","cited_arxiv_id":null,"evidence_quote":"Previous vertical jump work that supplies the landing-phase controller and the centroidal dynamics equations reused in the joint-space mapping."},{"cited_title":"Efficient dynamic computer simulation of robotic mechanisms,","cited_arxiv_id":null,"evidence_quote":"Defines composite rigid body inertia, which the paper invokes to justify treating inertia shaping as the only posture-control channel during flight."}],"review_version":1}