{"id":"601eadb3-cd30-48b5-9dfa-6555b20e1aa0","arxiv_id":"2412.16875","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"A planning and control framework for multi-axle swerve-drive robots that reduces swept area during turns by aligning vehicle heading with the path tangent and tracking the resulting trajectory with MPC.","lead":"Multi-axle robots that steer every wheel can reduce the space they sweep out during turns if the planner and controller are designed around that goal. This paper combines a signed-distance-field planner with model predictive control and tests it in simulation on a five-axle robot making a tight left turn.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is not supported because the MPC plant model (Eq. 22) and the wheel-velocity decomposition (Eq. 28) do not describe the actual rigid-body kinematics of a swerve-drive AMR; the reported simulation results may be self-consistent artifacts of the same erroneous model.","rationale":"The reader's verdict of REJECT is well supported, and I agree with the overall conclusion. However, the reader's declared weakest assumption is the swept-area proxy Jsv, while I see a more immediate and more decisive problem in the kinematics: Eq. (22) omits the rotation coupling between body-frame and global velocities, and Eq. (28) contains a sign error in the rotational contribution to longitudinal wheel velocity. These are internal inconsistencies independent of any modeling choice about swept area. If the plant model is not the kinematics of a real swerve-drive vehicle, then the reported tracking accuracy and swept-volume reduction can be simulation artifacts, regardless of whether Jsv faithfully represents swept area. For that reason my agreement is only partial: I would relocate the weakest link from the swept-area proxy to the kinematic correctness of the controller and wheel-command mapping. The concrete test is deliberately simple: a pure-rotation experiment distinguishes the correct and incorrect sign in Eq. (28), and a re-run with corrected MPC dynamics tests whether Table I is robust. This is not an ad hominem or a disagreement with the research direction; the problem is real and the framework may become viable after correcting the models, but the current evidence does not establish the central claim.","tokens_in":13693,"tokens_out":5170,"duration_ms":47065,"concrete_test":"Implement a pure-rotation consistency check in the swerve-drive simulator: command Vx=Vy=0 and a nonzero ω, then read back the resulting wheel velocity directions or steering angles. Rigid-body kinematics require V_ix = -ωYwi and V_iy = +ωXwi, while Eq. (28) predicts V_ix = +ωYwi. If the simulator matches the positive sign, the model is not a physical swerve-drive and Table I is an artifact of the wrong kinematics. If it matches the corrected sign, re-run the left-turn scenario with the corrected MPC rotation coupling in Eq. (22), holding all weights and scenario parameters fixed, and compare the resulting Sexcess and tracking errors against the reported values in Table I; any material change would confirm the published numbers depend on the erroneous model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the central claim is that the planner/MPC/wheel-command chain models the actual kinematics of a multi-axle swerve-drive AMR. This condition fails on two concrete internal points. First, the MPC plant model in Eq. (22), X(k+1)=A X(k)+B u(k) with A=I and B=T·I, treats the stated body-frame control vector u=[Vx,Vy,ω]^T as if it directly increments the global position [x,y] with no dependence on heading φ. If Vx,Vy are body-frame velocities, the correct discretized kinematics are x(k+1)=x(k)+T(Vx cosφ - Vy sinφ), y(k+1)=y(k)+T(Vx sinφ + Vy cosφ), φ(k+1)=φ(k)+Tω. If instead Vx,Vy are intended as global velocities, then the wheel decomposition in Eq. (28) is inconsistent, because it adds rotational terms to global velocities as though they were body-frame components. Second, Eq. (28) gives V_ix = Vx + ωYwi and V_iy = Vy + ωXwi. For a point at body-frame coordinates (Xwi,Ywi) with counterclockwise-positive yaw, the rigid-body rotational contribution is (-ωYwi, +ωXwi), so the top-row sign in Eq. (28) is opposite the standard convention. A simulator that implements the same incorrect kinematics can produce smooth, self-consistent trajectories, so the small tracking errors and the Sexcess values in Table I do not by themselves establish that the method works on a real multi-axle swerve-drive AMR. The swept-area proxy concern raised by the reader is real, but the kinematic error is more foundational: even a perfectly faithful swept-area objective would not rescue a controller whose state update and wheel-command mapping are not those of the physical vehicle.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified framework for multi-axle swerve-drive automated mobile robots that combines signed-distance-field (SDF) based trajectory planning with model predictive control (MPC) and per-wheel steering-angle conversion, with the goal of minimizing swept volume during turns. The planner first generates an A* path, smooths it with MINCO polynomials, and then optimizes a cost that includes obstacle distance, energy, time, and a heading-velocity alignment term intended to reduce swept area. The tracking layer uses a linear MPC with state X=[x,y,φ]^T and control u=[Vx,Vy,ω]^T, followed by a conversion to individual wheel speeds and steering angles. The method is evaluated in a Gazebo simulation of a 5-axle vehicle making a left turn, reporting an excess swept area of 23.14 m² versus 48.37 m² for the closest baseline, along with small tracking errors. The paper claims this is the first comprehensive approach to combine swept-volume minimization with independent axle control.","tokens_in":14057,"tokens_out":4363,"duration_ms":43409,"significance":"If the kinematic models and swept-area objective were correct, the paper would address a practically important problem for large multi-axle vehicles in constrained logistics environments, and the use of CUDA-accelerated swept-volume SDF evaluation plus an MPC tracking layer is a plausible pipeline. The authors also state an intent to open-source the work, which is commendable. However, the significance is currently undermined by load-bearing errors in the vehicle kinematics used for both MPC prediction and wheel command generation, and by the lack of a direct relationship between the optimized heading-velocity cost and the actual swept area. These issues mean the reported simulation improvements are not trustworthy evidence that the method would work on a real vehicle.","major_comments":[{"comment":"The discrete-time plant model X(k+1)=A X(k)+B u(k) with A=I, B=T·I, and u=[Vx,Vy,ω]^T is not a valid kinematic model for a rigid body. If Vx and Vy are body-frame velocities (as stated in Section III.A, where they are called longitudinal and lateral velocities), then the position update must depend on heading: x(k+1)=x(k)+T(Vx cosφ − Vy sinφ), y(k+1)=y(k)+T(Vx sinφ + Vy cosφ), φ(k+1)=φ(k)+Tω. If instead Vx and Vy are global-frame velocities, then the wheel decomposition in Eq. (28) is inconsistent because it adds body-frame rotational terms ωYwi and ωXwi to global components. Under either interpretation the MPC prediction does not describe the actual kinematics, so the small tracking errors in Table I could be artifacts of a simulator using the same incorrect model rather than evidence of real tracking performance.","section":"III.C, Eq. (22)"},{"comment":"The wheel-velocity decomposition contains a sign error in the transverse component. For a point fixed in the body frame at coordinates (Xwi, Ywi), with counterclockwise-positive yaw rate ω, the rigid-body velocity is Vix = Vx − ωYwi and Viy = Vy + ωXwi. Equation (28) instead gives Vix = Vx + ωYwi and Viy = Vy + ωXwi. This changes the computed wheel steering angles γi and speeds Vi in Eqs. (29)–(30), and if the same erroneous kinematics are implemented in the Gazebo simulation, the reported trajectory tracking and swept-area results do not demonstrate correct behavior on a physical swerve-drive AMR.","section":"III.C, Eq. (28)"},{"comment":"The swept-area cost Jsv is a heading-velocity alignment penalty, not the swept area S. The paper asserts that aligning the vehicle's long axis with the instantaneous velocity direction reduces swept area, but this is not established for a multi-axle vehicle whose rear axles off-track whenever yaw is nonzero. The optimizer never evaluates S during planning, so the claim that the planner 'minimizes swept volume' is not supported. A direct demonstration is needed, for example an ablation that compares the proposed proxy against an optimization that actually uses the swept-area SDF, or an analysis showing that the proxy bounds the true swept area. The post-hoc Sexcess metric in Section IV-C cannot substitute for a validated optimization objective.","section":"III.B.2, Eqs. (15)–(16)"},{"comment":"The experimental comparison conflates the planner and the controller. The Classic and Hierarchical baselines track the trajectory generated by the proposed planner but do not receive the swept-volume objective, so the Sexcess improvement in Table I may come primarily from the MPC tracking layer rather than from swept-volume-aware planning. The SVSDF baseline uses the proposed MPC but not the swept-area cost. To support the central claim, the comparison needs baselines that isolate the contribution of the Jsv term with a correct kinematic model, and the 'minimal swept area' used to define Sexcess must be precisely specified (for example, the convex hull of the vehicle's footprint, or the area of a straight corridor of vehicle width along the reference path).","section":"IV-C, Table I"}],"minor_comments":[{"comment":"The expression ∂Pj/∂Tj = Vj is dimensionally inconsistent as written: Pj is a position vector and Vj is a velocity, so the equality holds only under a specific time-scaling convention for the MINCO trajectory that should be stated explicitly.","section":"III.B.1, Eq. (4)"},{"comment":"The chain rule in Eq. (20) relies on the same derivative ∂Pj/∂Tj = Vj as Eq. (4), and therefore inherits the same need for a clearly defined time-scaling convention; as written, the terms ∂Jsv/∂Xj·VX,j and ∂Jsv/∂Yj·VY,j are also mixing position and velocity coordinates.","section":"III.B.2, Eq. (20)"},{"comment":"The tracking errors ey and eφ are reported only as ranges (e.g., ±0.04 m, ±0.03°), without the number of trials, the duration of each run, or the statistical variation, so it is unclear whether these represent worst-case, 1σ, or peak values over a single trajectory.","section":"IV-C, Table I and Fig. 6"},{"comment":"The statement 'A critical challenges is the swept volume' contains a grammatical error, and the terms 'swept volume' and 'swept area' are used interchangeably earlier in the text before the constant-height equivalence is introduced in Section III.A; this should be made consistent.","section":"I, Introduction"},{"comment":"The claim that the approach 'delivers life-saving improvements' is too strong for a simulation-only validation with an unverified kinematic model, and the promise of an open-source release should be accompanied by a functional repository link or a code availability statement in the manuscript.","section":"Abstract and V, Conclusion"}],"recommendation":"reject","confidential_remarks":"The two kinematic errors (Eq. (22) and Eq. (28)) are not presentation issues: they are load-bearing for the central claim that the planner/MPC/wheel-command chain controls a real multi-axle swerve-drive AMR. Since the simulation evidently uses the same erroneous equations, the favorable numbers in Table I may be self-consistent artifacts. Even if the sign in Eq. (28) were corrected, the missing rotation coupling in Eq. (22) would require a fundamentally different MPC formulation and a full re-evaluation. The paper also does not establish that the heading-velocity proxy in Eqs. (15)–(16) minimizes actual swept area, which is the stated core contribution. For these reasons I cannot recommend major revision as a proportionate outcome; the current manuscript's validity rests on modeling assumptions that appear to be incorrect."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The stress-test note is correct, and the reader is too. Eq. (22) and Eq. (28) do not describe the kinematics of a swerve-drive vehicle, and the paper's central claim is not supported as written.\n\nWhat is genuinely new and useful: the integrated pipeline is real. SDF-based MINCO planning, a heading-alignment cost as a swept-area proxy, MPC tracking, and per-wheel velocity-vector steering-angle conversion are combined in a way I have not seen in the cited literature. The CUDA implementation for swept-area SDF evaluation is a practical contribution, and the problem itself matters for logistics and confined-space maneuvering. Credit where it is due: the paper is clearly written, the prior-work connections are honest, and the open-source intent is good.\n\nThe soft spots are load-bearing. First, Eq. (22) treats body-frame velocities [Vx, Vy, omega] as if they directly increment the global state [x, y, phi]. The correct discretization needs the rotation terms involving phi. Since the simulator presumably uses the same wrong model, the tiny tracking errors in Table I are measures of self-consistency, not of physical tracking accuracy. Second, Eq. (28) has a sign error: for a point at (Xwi, Ywi) with counterclockwise-positive yaw, the rotational contribution to Vix is -omega*Ywi, not +omega*Ywi. This matters for a real vehicle's wheels. Third, the swept-area cost Jsv only penalizes the difference between heading and velocity direction; it never evaluates the swept area itself. The independent metric in Table I prevents circularity, but the optimizer may be minimizing a proxy that does not match the reported metric. Fourth, the experiments are a single noiseless simulation with no error bars, and the baselines are not full planning-plus-control pipelines. The 23.14 versus 48.37 comparison in Table I is suggestive, but not yet convincing.\n\nWho is this for? Researchers working on swept-volume-aware motion planning and multi-axle steering. The paper deserves a serious referee because the problem is real and the framework is plausible, but it needs major revision: fix the kinematics, re-run experiments with the corrected model, add multi-run statistics, compare against full-pipeline baselines, and either justify the heading proxy or optimize swept area directly. I would not cite it in its current form, but I would read a revised version.","headline":"Real problem, plausible pipeline, but the MPC plant model and wheel-velocity decomposition use the wrong kinematics, so the validation does not yet transfer to hardware.","tokens_in":14703,"tokens_out":2514,"would_cite":false,"duration_ms":24786,"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 swept-area-aware SDF planner plus MPC tracking cuts a five-axle AMR's excess swept area from 48.37 m² to 23.14 m² in a simulated left turn.","keywords":["swept volume minimization","multi-axle AMR","swerve-drive","model predictive control","signed distance field path planning","trajectory optimization","independent wheel steering","autonomous logistics"],"falsifier":"In the same Gazebo left-turn scene, rasterize the vehicle footprint at every pose along the planned and tracked trajectory, take the union to get the actual swept area, and compute S_excess; then rerun the second optimization stage with W_sv = 0. If removing the swept-area cost does not increase the measured excess area, or if trajectories with lower J_sv do not have smaller S_excess, the claimed mechanism is not what produces the reported reduction.","tokens_in":13370,"feed_emoji":"🚚","tokens_out":6722,"duration_ms":55912,"temperature":0.7,"pith_summary":"This paper is trying to show that the swept volume, the extra ground a long multi-axle vehicle occupies during a turn, can be actively minimized by coupling a swept-area-aware path planner with a model predictive controller that steers each wheel independently. The proposed pipeline plans a path with obstacle avoidance and a cost that keeps the vehicle's length axis aligned with its direction of travel, then tracks that path with MPC and converts the body-level velocity command into per-wheel steering angles. In a simulated left-turn scenario with a five-axle, 8.1-meter vehicle, the method reports an excess swept area of 23.14 m², about half the 48.37 m² of the best comparison method, along with 1.17 seconds of planning time and lateral and heading tracking errors within ±0.04 m and ±0.03°. If these results hold, warehouse and highway logistics vehicles could maneuver through tight spaces with smaller safety zones, reducing the risk of hitting pedestrians or obstacles. The paper also claims this is the first comprehensive approach to combine swept-volume minimization with independent axle control.","feed_headline":"Turning footprint shrinks to 23 m² with swept-area MPC","feed_subtitle":"An 8-meter five-axle logistics robot keeps its length along the path, cutting extra swept ground by half in a left-turn test.","key_machinery":"The load-bearing object is the swept-area cost Jsv = Σ (Δφ_j)², with Δφ_j = φ_j − atan(VY,j/VX,j). This is a heading-velocity alignment penalty: at each control point it drives the vehicle's heading φ_j to coincide with the direction of its velocity vector, so the long axis of the rectangular body stays tangent to the path and the vehicle sweeps a narrower corridor. It sits inside the second optimization stage alongside an SDF-based obstacle cost Job that penalizes proximity to obstacles using the implicit signed distance field of the rectangular footprint, and the whole trajectory is parameterized as MINCO, the minimum-control-effort polynomial trajectory class, so gradients with respect to control points and segment times are available. The MPC layer then takes the optimized body-level control (Vx, Vy, ω) and derives each wheel's steering angle γ_i and speed Vi through the rigid-body velocity relation, which is what turns a point-mass optimal control problem into commands for independently steerable axles.","core_discovery":"The central claim is that swept-volume minimization for multi-axle swerve-drive AMRs can be achieved by a two-stage optimization followed by MPC tracking: first smooth an A* path, then optimize it under SDF obstacle costs and a swept-area cost Jsv that penalizes the squared angular difference between the vehicle heading φ_j and the velocity direction atan(VY,j/VX,j), and finally track it with an MPC controller that converts the optimal body velocity (Vx, Vy, ω) into per-wheel steering angles γ_i = arctan(Viy/Vix). In their Gazebo left-turn scenario, this reduces excess swept area to 23.14 m² versus 48.37 m² for the hardest baseline, with planning time of 1.17 seconds and tracking errors within ±0.04 m and ±0.03°. The paper presents this as the first comprehensive approach to combine these elements.","pith_inferences":["For a fixed chassis, the benefit should scale with vehicle length and the number of steerable axles; a short single-axle robot would likely show little difference, so the method is most relevant to long multi-axle platforms.","The heading-alignment idea could transfer to articulated vehicles by imposing a per-trailer alignment cost, though the paper does not test this.","Because the MPC tracker is separable from the planner, the same per-wheel steering conversion could be applied to any path planner, not just the SDF-based one used here.","Pairing the swept-area proxy with a swept-region SDF, so that the planner checks collisions against the area the vehicle will actually occupy, could remove the need to trust the proxy; this would be a natural follow-up."],"forward_implications":["In the simulated left-turn scenario, the proposed pipeline reduces excess swept area from 48.37 m² for the best baseline to 23.14 m², roughly halving the extra ground covered.","Planning time of 1.17 seconds with CUDA-accelerated swept-volume SDF estimation is fast enough for near-real-time replanning in a static scenario.","The MPC tracker holds lateral error within ±0.04 m and heading error within ±0.03° on the planned trajectory, so the theoretical swept-area reduction is not lost to tracking deviations.","Computing per-wheel steering angles from the velocity vector allows a single point-mass MPC controller to command all independently steerable wheels, removing the need for a separate steering-center mode switch.","SVSDF-style planning can be combined with this MPC tracker and still achieve accurate tracking, indicating the tracking component generalizes beyond the paper's own planner."],"supporting_citations":[{"why":"Supplies the classic tractor-trailer tracking baseline the proposed method is compared against.","marker":"[1]"},{"why":"Supplies the continuous implicit SDF optimization method for arbitrary robot shapes used for the obstacle cost.","marker":"[4]"},{"why":"Supplies the swept-volume SDF baseline SVSDF and the implicit swept-volume representation the paper extends.","marker":"[5]"},{"why":"Supplies the hierarchical multi-mode steering baseline and the idea of velocity-vector steering modes.","marker":"[65]"},{"why":"Supplies the MINCO polynomial trajectory representation that parameterizes both optimization stages.","marker":"[77]"},{"why":"Supplies the L-BFGS solver used in the first smoothing optimization stage.","marker":"[78]"},{"why":"Supplies the LMBM solver used in the second swept-volume-aware optimization stage.","marker":"[79]"},{"why":"Supplies the Armijo line search used to find the closest pose when estimating swept-area SDF over a grid.","marker":"[80]"}],"fun_headline_variants":["Swept area halved: 23 m² vs 48 m² in left-turn test","MPC tracking cuts AMR swept volume to 23 m²","Real-time SDF path planning shrinks turning footprint","Multi-axle AMR: swept-volume-aware MPC cuts space by half"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire swept-volume reduction rests on the assumption that penalizing the squared angle between the vehicle's heading and its velocity direction is a faithful proxy for minimizing the actual swept area, since the optimizer never directly computes or evaluates that area.","fun_headline_variants_meta":{"raw":{"variants":["Swept area halved: 23 m² vs 48 m² in left-turn test","MPC tracking cuts AMR swept volume to 23 m²","Real-time SDF path planning shrinks turning footprint","Multi-axle AMR: swept-volume-aware MPC cuts space by half"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000985,"raw_usage":{"total_tokens":4198,"prompt_tokens":987,"completion_tokens":3211,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":3139}},"tokens_in":603,"tokens_out":3211,"duration_ms":21838,"temperature":1.0,"reasoning_tokens":3139,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T06:01:37.548379+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the same Gazebo left-turn scene, rasterize the vehicle footprint at every pose along the planned and tracked trajectory, take the union to get the actual swept area, and compute S_excess; then rerun the second optimization stage with W_sv = 0. If removing the swept-area cost does not increase the measured excess area, or if trajectories with lower J_sv do not have smaller S_excess, the claimed mechanism is not what produces the reported reduction.","supporting_citations":[{"cited_title":"Modelling and dynamic tracking control of industrial vehicles with tractor-trailer structure,","cited_arxiv_id":null,"evidence_quote":"Supplies the classic tractor-trailer tracking baseline the proposed method is compared against."},{"cited_title":"Implicit swept volume sdf: Enabling continuous collision- free trajectory generation for arbitrary shapes,","cited_arxiv_id":null,"evidence_quote":"Supplies the swept-volume SDF baseline SVSDF and the implicit swept-volume representation the paper extends."},{"cited_title":"Hierarchical control strategies for multi-mode steering system of emergency rescue vehicle,","cited_arxiv_id":null,"evidence_quote":"Supplies the hierarchical multi-mode steering baseline and the idea of velocity-vector steering modes."},{"cited_title":"lbfgs: Efficient l-bfgs and owl-qn optimization in r,","cited_arxiv_id":null,"evidence_quote":"Supplies the L-BFGS solver used in the first smoothing optimization stage."},{"cited_title":"Limited memory bundle method and its variations for large-scale nonsmooth optimization,","cited_arxiv_id":null,"evidence_quote":"Supplies the LMBM solver used in the second swept-volume-aware optimization stage."},{"cited_title":"Minimization of functions having lipschitz continuous first partial derivatives,","cited_arxiv_id":null,"evidence_quote":"Supplies the Armijo line search used to find the closest pose when estimating swept-area SDF over a grid."}],"review_version":1}