{"id":"4d471cba-0e0d-4fdb-b469-359f335b4d71","arxiv_id":"2505.20926","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A movable center-of-mass battery plus a hybrid Fuzzy-PID and VUFC-ADRC control scheme improves steady-state steering and walking stability of a deformable wheel-legged vehicle.","lead":"This paper equips a deformable wheel-legged vehicle with a sliding battery that shifts the center of mass, and it designs a hierarchical controller to improve stability during both driving and walking. A reader might care because the approach addresses the unusual problem of stabilizing one robot in two very different motion modes using a single hardware adjustment.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Walking-stability claim rests on an unvalidated mapping from simulation-trained K-means cluster labels to real-robot COP data; no no-control hardware baseline is reported.","rationale":"The reader's weakest_assumption and my independent reading converge on the same concern: the K-means stability labels are trained purely on simulation and the paper self-reports a one-level hardware mismatch in Section 6.3, yet the hardware walking claim is made entirely in terms of those labels and there is no no-control hardware baseline. This is not merely a disagreement with the community consensus; it is an internal validation gap in the paper's own reported evidence. The vehicular-state steering claim is comparatively better supported: it uses a standard fixed-steering-angle test with a before/after comparison, and the improvement from K=0.00097 toward K_d=0.0024 is backed by measured yaw rate and slip-angle data. The walking claim, however, has no comparable before/after. A no-control walking baseline is the minimal, decisive experiment: it determines whether the reported level-2 hardware stability is due to the controller or simply to the nominal gait being stable. Machine-checked proofs are not present, and the 8 hand-tuned parameters (fuzzy-PID gains, ADRC bandwidths, b0 values) are not derived from the model, which further raises the risk that the simulation results are tuned, but the primary logical gap remains the unvalidated label mapping and missing baseline. I therefore keep the reader's CONDITIONAL verdict: the claim is plausible and partially supported, but acceptance of the walking-stability contribution should be contingent on the no-control hardware baseline and on demonstrating that the simulation-trained cluster labels are meaningful for real-robot COP data.","tokens_in":24217,"tokens_out":1477,"duration_ms":14134,"concrete_test":"Re-run the walking test on the physical prototype with the COM adjustment mechanism disabled (sliders locked at nominal position) under identical gait and terrain, record the same COP-derived ZMP time series, and classify it using the Table 1 cluster centers. If the no-control hardware ZMP deviations fall in the same or lower levels (e.g., also level 2) as the controlled run, the claimed walking-stability improvement is not supported; if they degrade to level 3 or worse, the claim is supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central walking-stability claim is that VUFC-ADRC with K-means grading keeps real walking at stability level 2 (Section 6.3). The load-bearing link is that the cluster centers in Table 1, trained entirely on ADAMS simulation data for the nominal gait (Section 4.2), remain valid labels for the real robot's ZMP/COP measurements. The paper itself reports a systematic one-level degradation in hardware (Section 6.3), which shows the simulation-to-reality mapping is already miscalibrated. The stability 'level' therefore is not a measured physical quantity but an artifact of cluster boundaries learned in simulation; if the real-robot COP distribution differs more substantially in other gait phases or under disturbances, the level labels used to trigger the variable-universe scaling factors will be systematically wrong, and the claimed 'level 2' result cannot be interpreted as a stability margin. Furthermore, no hardware experiment without the COM controller is reported for walking, so the improvement claim has no baseline against real disturbances; the comparison to uncontrolled simulation does not establish the hardware effect.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a center-of-mass (COM) adjustment mechanism for a reconfigurable wheel-legged robot and a hierarchical control strategy for two configurations. The upper-level hybrid automaton switches between a vehicular steady-state steering controller and a humanoid walking controller. The steering controller uses Fuzzy-PID to move an X-direction slider so that the vehicle stability factor K tracks a desired understeering value K_d = 0.0024, improving the simulated and measured K from 0.00097. The walking controller uses a variable-universe fuzzy controller combined with active disturbance rejection control (VUFC-ADRC), with a K-means clustering stage that grades the walking stability level from ZMP deviation and deviation rate. Simulation and hardware tests show the ZMP tracked near the planned reference and stability levels mostly at levels 1 and 2 in simulation and at level 2 in the hardware test; the paper also reports that hardware levels are one level higher than simulation. The central claim is that slider control significantly improves multi-configuration motion stability.","tokens_in":24523,"tokens_out":8452,"duration_ms":85558,"significance":"The idea of using a dedicated COM actuator for both vehicle steering and walking stability is useful, and the steering result is the strongest part: the stability factor is a standard objective, the before/after K values are external to the controller, and the hardware curves in Fig. 12 show a consistent understeering improvement. The walking control architecture is more speculative. The K-means-graded VUFC-ADRC is an interesting combination, but the hardware validation currently does not isolate the effect of the COM control from the gait planner, and the stability-level labels are not calibrated to a physical margin. The paper is transparent about the one-level discrepancy between simulation and test in Section 6.3, which is a useful limitation but also highlights the need for further validation. If the authors add a hardware baseline and validate the K-means labels against a physical ZMP margin, the contribution would be solid.","major_comments":[{"comment":"The central walking-stability improvement claim is not supported on hardware: no no-control or alternative-controller baseline is reported for the walking test. The hardware evidence in Figs. 13(a)-(b) is that the controlled ZMP tracks the ideal trajectory, and Figs. 13(c)-(d) show the resulting stability levels. Tracking a planned reference does not by itself demonstrate that the COM controller improves stability; the comparison against the PID in the simulation of Section 5.2 is not an experimental baseline. Please add a hardware baseline with the COM slider fixed, or with PID-only control, or otherwise restrict the hardware claim to 'stability is maintained' rather than 'significantly improved.'","section":"6.3 and Fig. 13"},{"comment":"The K-means cluster centers in Table 1 are trained exclusively on ADAMS simulation data for the nominal gait, yet they are applied to real-robot COP measurements in Section 6.3 to label the stability level. The paper itself reports that the hardware stability levels are systematically one level higher than the simulation results (Section 6.3), demonstrating a simulation-to-hardware calibration gap. Consequently, the reported 'level 2' result is a cluster index with unvalidated physical meaning; if the real COP distribution differs beyond the observed shift, the variable-universe scaling factors will be triggered by mislabeled levels and the claimed walking improvement may not transfer. Furthermore, the level is defined from the same ZMP error signal that the controller minimizes, so the level improvement is partly self-referential unless calibrated to a physical stability margin. Please validate the cluster labels against a physical margin (e.g., distance from ZMP to the support-polygon boundary) on hardware, or retrain/calibrate the clusters with hardware data and show that the labels remain consistent.","section":"4.2, Table 1, 6.3"},{"comment":"The homogeneous transformation matrix in Eq. (9) is not a valid transformation matrix as printed: entries such as 'cosθ_iz cosθ_ix - cosθ_iz cosθ_ix' and the placement of squared cosine terms are internally inconsistent, and the matrix does not have the standard rigid-body form. Since Eqs. (11)-(15) compute COM and ZMP from this transformation, the kinematic model used for the walking controller and simulation should be corrected or replaced with a correctly rendered version, or the derivation should be referenced to a reliable source.","section":"3.2, Eq. (9)"}],"minor_comments":[{"comment":"Equation (6) appears twice verbatim in the text; one duplicate should be removed.","section":"3.1"},{"comment":"Equation (17) states '105' actual ZMP samples, but the surrounding text says 10,000 steps with 10 samples per step, which is 100,000 samples; the exponent formatting should be corrected and the notation made explicit.","section":"4.2, Eq. (17)"},{"comment":"Reference [28] is cited for the ZMP/COP approximation and reference [29] for the fixed front wheel steering angle method, but the reference list identifies [28] as an automobile handling test standard and [29] as a trajectory optimization paper; the citations should be remapped.","section":"6.1, 6.2"},{"comment":"Equation (25) places the observer poles at -ω0 while Eq. (26) and the simulation parameters in Section 5.2 use ω_e; unify the notation for the observer bandwidth.","section":"4.2 c)"},{"comment":"The simulation claims 'anti-disturbance ability' but does not specify what external disturbance was applied. Please state the disturbance model and magnitude, otherwise the comparison between the VUFC-ADRC variants is not reproducible.","section":"5.2"},{"comment":"The weighted distance in Eq. (18) uses fixed weights 0.7/0.3. A brief justification of these weights, or a sensitivity study of the resulting cluster centers to them, would strengthen the K-means grading stage.","section":"4.2, Eq. (18)"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the manuscript's novelty is incremental relative to the group's prior work [1]-[6]; the main new ingredient is the K-means-graded VUFC-ADRC walking controller. The steering part is publishable, but the walking-stability claim needs a hardware baseline and a validation of the simulation-trained cluster labels before acceptance. I recommend major revision rather than reject, since the identified issues are addressable within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this paper has a genuine hardware result in the steering half and a load-bearing gap in the walking half. The steering test shows the stability factor moving from 0.00097 to near the target 0.0024 with the COM slider, and the before/after yaw-rate curves look consistent. That is a legitimate demonstration, and it is the paper's strongest contribution.\n\nWhat is new is narrow: a two-axis battery-as-slider COM adjustment mechanism integrated with a hybrid-automata mode switch, Fuzzy-PID for steering, and VUFC-ADRC plus K-means stability grading for walking. None of the ingredients is new, and the ADRC derivation is textbook. The novelty is the integration on this specific deformable wheel-legged platform, and that is fine for an engineering paper.\n\nThe soft spots are concentrated in the walking claim. The K-means cluster centers in Table 1 are trained exclusively on ADAMS simulation of the nominal gait, then applied to real-robot COP data. The paper itself reports a one-level degradation from simulation to hardware, so the mapping is already miscalibrated. That does not make the walking control worthless—the test shows the COP roughly tracking the ideal ZMP—but it means the 'level 2' in the test is a sim-derived label, not a measured stability margin. Bigger problem: there is no no-control hardware baseline for walking. The simulation comparison against PID and VUFC-ADRC without K-means is fine, but it does not tell us what the real robot does without the COM controller. So the claim that the slider significantly improves walking stability on hardware is under-supported.\n\nOther nits: there are typos (Eq. (6) is duplicated), the fuzzy rules are hand-tuned with no sensitivity analysis, and the number of free parameters in ADRC and fuzzy controllers is large. None of this is disqualifying, but a reviewer should push on how sensitive the results are to those choices.\n\nBottom line: the steering result deserves publication, and the hardware platform is real. The walking section needs a no-control baseline and a better justification or re-calibration of the K-means labels before the central claim is accepted. I would send this to peer review, not desk reject, but with a request for revision focused on the walking evidence.","headline":"A real hardware steering result and a walking claim that needs a baseline before it is believed.","tokens_in":25008,"tokens_out":2662,"would_cite":false,"duration_ms":28139,"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 sliding battery keeps a transforming robot stable in both modes.","keywords":["unmanned deformable vehicle","center-of-mass adjustment mechanism","stability factor","zero-moment point","K-means clustering","variable universe fuzzy control","active disturbance rejection control","hybrid automata control"],"falsifier":"Freeze the COM sliders at their center positions and rerun the steady-state steering and walking tests under the same disturbances; if the stability factor still holds near $0.0024$ and the ZMP still stays at level 2, the mechanism is not doing the work. Sharper still: recompute the K-means cluster centers from hardware-measured (ZMP error, ZMP error-rate) samples during real walking and check whether they coincide with Table 1; mismatch would mean the stability-level labels that scale the fuzzy universe are miscalibrated.","tokens_in":24054,"feed_emoji":"🤖","tokens_out":6771,"duration_ms":63806,"temperature":0.7,"pith_summary":"The paper claims that one physical mechanism -- a two-axis slider that shifts the vehicle's center of mass -- can actively stabilize both motion configurations of a wheel-legged vehicle that transforms between a car and a biped. In the vehicular state, moving the X-slider changes the front and rear wheelbases, which changes the steering stability factor $K$; a Fuzzy-PID controller drives $K$ to the understeering target $0.0024$ from an uncontrolled $0.00097$. In the humanoid state, moving X- and Y-sliders keeps the zero-moment point (ZMP) on its desired trajectory; a K-means-graded variable-universe fuzzy controller with active disturbance rejection (VUFC-ADRC) keeps walking stability mostly at levels 1 and 2 in simulation and at level 2 in physical tests. The reason to care is that stability is achieved by a dedicated lightweight actuator, so the leg joints can keep executing the planned gait and the steering system does not have to interrupt its motion to recover from disturbances.","feed_headline":"A sliding battery keeps a transforming robot stable","feed_subtitle":"COM-slider control holds the steering stability factor near 0.0024 and walking stability at level 2.","key_machinery":"The carrying object is the center-of-mass adjustment mechanism: a T-shaped pair of X- and Y-direction slides mounted on the body, with the battery itself serving as the movable mass, driven by a ball screw and modeled as the linear second-order electromechanical system $\\ddot{y} = -\\frac{B_z}{J}\\dot{y} + \\frac{K_a K_t r_g}{J}u$. The two identities that carry the argument are the steering stability factor $K = \\frac{m}{L^2}\\left(\\frac{a}{k_2}-\\frac{b}{k_1}\\right)$ and the zero-moment point (ZMP); slider displacement enters both. The new coupling is K-means clustering on (ZMP error, ZMP error-rate) data, which turns the qualitative ZMP criterion into five discrete stability levels $L_1$-$L_5$, and a variable-universe fuzzy controller that expands or contracts its input universe according to the current level, followed by an ADRC that tracks the commanded slider displacement.","core_discovery":"The paper's central discovery is that center-of-mass adjustment alone is a sufficient control channel for both steady-state steering stability and walking stability. For steering, the problem is compressed into the stability factor identity $K = \\frac{m}{L^2}\\left(\\frac{a}{k_2} - \\frac{b}{k_1}\\right)$: shifting the X-slider changes $a$ and $b$, and closed-loop Fuzzy-PID control converges $K$ to the desired value $0.0024$, improving overshoot from $14.99\\%$ to $9.19\\%$, rise time from $1.05\\,\\mathrm{s}$ to $0.35\\,\\mathrm{s}$, and settling time from $2.37\\,\\mathrm{s}$ to $1.52\\,\\mathrm{s}$ in simulation. For walking, the paper converts the qualitative ZMP stability criterion into a quantitative five-level scale by running K-means clustering on $10^5$ ADAMS-simulated ZMP deviation samples, then uses the stability level to scale the universe of a variable-universe fuzzy controller whose output is a desired slider displacement; an active disturbance rejection controller tracks that displacement. The reported results are a reduction in maximum ZMP tracking error from $0.037\\,\\mathrm{m}$ under PID to near the ideal trajectory, and, with K-means grading, a $14.86\\%$ reduction in overshoot and $3.6\\%$ reduction in steady-state error relative to VUFC-ADRC without grading.","pith_inferences":["The one-level gap between simulated and hardware stability levels suggests a systematic calibration offset in the ZMP model; if the gap grows with terrain or gait, the cluster centers would need to be retrained per operating condition rather than once.","The constant target $K_d = 0.0024$ is a single operating point; a speed- or load-dependent target would be a natural extension and would test whether the slider authority is adequate at the characteristic speed of 75 km/h.","The mechanism's authority is bounded by slider travel, so for large disturbances the next step would be blending slider motion with leg or steering actuation; the hierarchical architecture already has the discrete states to host such a fallback."],"forward_implications":["A conventional fixed-COM vehicle cannot change its stability factor; the mechanism gives the unmanned deformable vehicle an actively tunable understeering capability that can be re-targeted for different loading and speed conditions.","Because stability correction happens through the COM slider rather than through leg joints or steering, the planned gait and steering can continue undisturbed while the controller recovers stability.","The K-means five-level grading turns the ZMP criterion from a binary safe/unsafe check into a graded safety margin that can be monitored in real time and acted on before the ZMP leaves the support region.","The hybrid automata upper controller gives a formal way to switch the same physical actuator between the two configurations, so adding future motion modes only requires adding new discrete states rather than new hardware."],"supporting_citations":[{"why":"Supplies the stability factor $K$ and yaw-rate-gain definitions that the vehicular controller regulates.","marker":"[18]"},{"why":"Provides the understeering target value $K_d = 0.0024$ used as the control objective in the Fuzzy-PID loop.","marker":"[27]"},{"why":"Supplies the ZMP-based biped walking stability criterion that the paper quantifies and grades.","marker":"[7]"},{"why":"Defines ZMP support areas under friction, the qualitative stability boundary that the five-level grading refines.","marker":"[8]"},{"why":"Provides the inverted-pendulum gait planning model used to generate the humanoid walking trajectory.","marker":"[19]"},{"why":"Demonstrates clustering-based stability discrimination for vehicles, the template for K-means ZMP grading.","marker":"[22]"},{"why":"Grounds the ADRC extended-state-observer eigenvalue placement used to stabilize the slider tracking loop.","marker":"[25]"}],"fun_headline_variants":["Sliding mass steadies a robot that transforms from vehicle to biped","COM slider control improves stability in both driving and walking modes","Center-of-mass slider keeps transformable robot stable on wheels and legs","Fuzzy-controlled slider balances a shape-shifting robot's drive and walk","One slider controls stability for both rolling and walking of a deformable bot"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The walking-stability grading rests on the assumption that the real robot's ZMP deviations fall into the same clusters as the ADAMS simulation of the nominal gait, so the Table 1 cluster centers remain valid labels on hardware; the paper's own test already shows the real stability level running one level higher than simulation.","fun_headline_variants_meta":{"raw":{"variants":["Sliding mass steadies a robot that transforms from vehicle to biped","COM slider control improves stability in both driving and walking modes","Center-of-mass slider keeps transformable robot stable on wheels and legs","Fuzzy-controlled slider balances a shape-shifting robot's drive and walk","One slider controls stability for both rolling and walking of a deformable bot"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001071,"raw_usage":{"total_tokens":4501,"prompt_tokens":976,"completion_tokens":3525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":3433}},"tokens_in":592,"tokens_out":3525,"duration_ms":28255,"temperature":1.0,"reasoning_tokens":3433,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:43:51.334971+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Freeze the COM sliders at their center positions and rerun the steady-state steering and walking tests under the same disturbances; if the stability factor still holds near $0.0024$ and the ZMP still stays at level 2, the mechanism is not doing the work. Sharper still: recompute the K-means cluster centers from hardware-measured (ZMP error, ZMP error-rate) samples during real walking and check whether they coincide with Table 1; mismatch would mean the stability-level labels that scale the fuzzy universe are miscalibrated.","supporting_citations":[{"cited_title":"Automobile Theory,","cited_arxiv_id":null,"evidence_quote":"Supplies the stability factor $K$ and yaw-rate-gain definitions that the vehicular controller regulates."},{"cited_title":"Vehicle stability,","cited_arxiv_id":null,"evidence_quote":"Provides the understeering target value $K_d = 0.0024$ used as the control objective in the Fuzzy-PID loop."},{"cited_title":"ZMP Support Areas for Multicontact Mobility Under Frictional Constraints,","cited_arxiv_id":null,"evidence_quote":"Defines ZMP support areas under friction, the qualitative stability boundary that the five-level grading refines."},{"cited_title":"Combinatorial clustering based lateral stability discrimination method for intelligent vehicles,","cited_arxiv_id":null,"evidence_quote":"Demonstrates clustering-based stability discrimination for vehicles, the template for K-means ZMP grading."},{"cited_title":"Modern Control Theory,","cited_arxiv_id":null,"evidence_quote":"Grounds the ADRC extended-state-observer eigenvalue placement used to stabilize the slider tracking loop."}],"review_version":1}