{"id":"157a5ca4-6dd1-4694-80ef-5596d584468e","arxiv_id":"2507.12578","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A deep bilinear Koopman model with a cumulative error regulator improved trajectory-tracking accuracy in vehicle control, reducing tracking error by 33% to 68% versus baselines in hardware-in-the-loop tests.","lead":"The authors train a deep neural network to learn a bilinear Koopman model of car dynamics and embed it in a model predictive controller with an integral-error module. In hardware-in-the-loop tests with a high-fidelity car simulator, their controller tracks a double lane change more tightly than three baseline controllers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (40) freezes the lifted state for the whole horizon and also drops the curvature bilinear term H3(Z_t * kappa); the MPC is therefore not solving the learned bilinear model, so the closed-loop advantage over CER-MDK may not demonstrate bilinear-model benefit.","rationale":"The reader's weakest assumption points to the same Section III-H linearization, and that is indeed the most load-bearing spot for the closed-loop claim. I sharpen it with a concrete omission: even the local linearization is incomplete because the curvature bilinear channel H3 is dropped from Eq. (40). This matters because Section III-F identifies H3 as contributing to the Frenet tracking states that drive the headline ey/e_psi improvements. The concern is testable and addressable, but it does not by itself overturn the paper. The HIL deployment, the 0.58 ms measured turnaround, and the consistent open-loop prediction results are independent evidence that the overall pipeline is real-time-capable and that the learned model has predictive value. What is not established is the attribution of the closed-loop advantage to the bilinear dynamics as opposed to the frozen-state linearization schedule plus CER. Since the paper already presents this as an approximation and the authors could settle the issue with a corrected or exact-bilinear comparison, the appropriate disposition remains the reader's CONDITIONAL verdict; no change is required.","tokens_in":18500,"tokens_out":11324,"duration_ms":128747,"concrete_test":"Re-run the Section IV DLC closed-loop test with two modifications, keeping all weights, horizons, and baselines unchanged: (i) include the omitted affine term H3(Z_t * kappa) in Eq. (40); (ii) replace the frozen-Z_t recurrence with the exact bilinear recurrence (24) as the MPC predictor, using fixed-point iteration or a nonlinear solver. Compare RMSE of S, ey, and e_psi with Table III and record max ||Z_k - Z_t|| / ||Z_t|| over the prediction horizon. If the corrected-linear or exact-bilinear controller changes ey RMSE by more than 10%, or if the max lift-state drift exceeds the neighborhood in which Eq. (40) matches Eq. (24), the central claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the learned bilinear Koopman model is responsible for the tracking improvements in Table III rests on the MPC internal model in Eq. (40). There, the bilinear term is replaced by a term frozen at the initial lifted state Z_t, so within the 20-step horizon the recurrence is linear in (Z, u). The model therefore does not propagate the multiplicative coupling U_i^k * Z_k that defines the bilinear representation (24); it only schedules an input matrix once per step from the current lifted state. The paper gives no quantitative bound on how far Z_k drifts from Z_t over the horizon, and the single DLC scenario with constant kappa = 0.001 cannot establish that the frozen-state neighborhood is valid in the maneuvers where the method claims an advantage. More concretely, Eq. (40) also omits the curvature-channel bilinear term H3(Z_t * kappa): the derivation moves from sum_{i=1}^m H_i(Z_t u_i^k) to (B_u + sum_j hatH_j[Z_{t,j}]) u_k, with m = 2, and never adds the known affine term H3(Z_t * kappa) that a true linearization of Eq. (24) would include. Thus even at the first prediction step the MPC's internal model is not the local linearization of the learned bilinear model. If including that term or using an exact-bilinear predictor changes the reported ey/e_psi RMSE materially, the claimed bilinear advantage is partly an artifact of the MPC formulation rather than of modeling fidelity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Multi-Step Deep Bilinear Koopman Network (MDBK-Net) that jointly learns lifting functions and bilinear Koopman dynamics for vehicle state evolution in the Frenet frame, and integrates the learned model with a Cumulative Error Regulator (CER) into a model predictive controller (CER-MDBK-MPC). The model is evaluated in open-loop prediction against EDMDK and the authors' previous linear MDK-Net on 1,425 CarSim test trajectories, and in closed-loop hardware-in-the-loop experiments on a dSPACE SCALEXIO system with a CarSim RT vehicle model. The paper reports lower prediction RMSE than the baselines and closed-loop tracking-error reductions of 33% versus CER-MDK-MPC, 38% versus EDMDK-MPC, and 68% versus LTI-MPC, with an average controller compute time of 0.58 ms per step.","tokens_in":18891,"tokens_out":6746,"duration_ms":76760,"significance":"If the claims hold, the work is a useful demonstration of data-driven Koopman modeling for vehicle control: it shows that learned bilinear couplings between inputs and lifted states can be trained with a multi-step loss and deployed in real-time MPC on embedded hardware. The open-loop evaluation over 1,425 trajectories is a strength, as is the HIL validation on a commercial real-time simulator. The comparison against a linear Koopman baseline with the same CER module is a reasonable attempt to isolate the effect of bilinear modeling. However, the closed-loop evidence is limited to a single scenario, and the MPC internal model is not exactly the learned bilinear model, which weakens the attribution of the reported improvements to bilinear modeling.","major_comments":[{"comment":"The predictive model actually implemented in the MPC is not the learned bilinear dynamics in Eq. (24). First, the equality constraint in (39b) sums only over i = 1..m, omitting the curvature-channel bilinear term H_3(Z_k * κ) that appears in the learned dynamics; even when κ is constant, H_3(Z_t κ) is an affine offset that is discarded. Second, Eq. (40) fixes the lifted state to its initial value Z_t throughout the entire prediction horizon, so the resulting recurrence is linear in the input with a state-scheduled input matrix. The paper attributes the Table III tracking improvements to the 'bilinear terms introduced in the proposed framework,' but the MPC is not solving the learned bilinear model even at the first prediction step. The comparison against CER-MDK-MPC therefore contrasts two different learned models plus a gain-scheduling approximation, not a controlled test of bilinear modeling. Please either include the H_3 term in the MPC formulation, report the drift ||Z_k - Z_t|| over the 20-step horizon for the test scenario to justify the freezing, or explicitly discuss why this approximation is valid and what its error magnitude is.","section":"III-H, Eqs. (39b) and (40)"},{"comment":"The closed-loop validation rests on a single double-lane-change scenario, and the text reports a constant curvature profile of κ = 0.001 m⁻¹, which is not representative of an actual lane-change maneuver (a real DLC has varying curvature). No repeated runs, variations in initial conditions, speeds, or road geometries are provided, so the reported RMSE reductions (e.g., 97% for e_y versus EDMDK) are based on one trajectory. The abstract's claim of 'significant reductions' is therefore not statistically substantiated. Please evaluate the controllers on multiple scenarios (varying curvature profiles, speeds, and maneuvers), report mean and standard deviation over repeated runs, and, if possible, include a significance test. This is especially important because the closed-loop controller uses the frozen-state approximation, and its validity may depend on the maneuver.","section":"IV, Table III and Fig. 7"}],"minor_comments":[{"comment":"The RMSE values are described as being computed after normalizing each state using the mean and standard deviation of the test dataset, yet the table headers show physical units (km/h, deg/s, m). This is potentially misleading; please report physical units with normalized values clearly distinguished, or provide both normalized and unnormalized errors.","section":"IV, Table II"},{"comment":"The sentence claiming reductions 'by 5%, 97%, and 88% compared to the EDMDK-MPC' is numerically inconsistent with Table III: for S the reduction is approximately 2.4% (0.206 vs 0.211), not 5%. Please correct this value.","section":"IV, text after Table III"},{"comment":"The statement 'The same optimal control problem formulation is employed for all controllers' is imprecise: the dynamic constraints differ between the baselines and the proposed method. Please clarify that the cost weights, constraints, and prediction horizon are identical, while the internal models differ.","section":"IV, experimental setup"},{"comment":"The notation H_i = [h_{i1}, ..., h_{ip}] and \\hat{H}_j = [h_{1j}, ..., h_{mj}] is not fully defined; if h_{ij} are columns of H_i (each of dimension p), then \\hat{H}_j is p × m, but this should be stated explicitly to avoid confusion.","section":"III-H, Eq. (40)"},{"comment":"The paper alternates between 'MDK-MPC' and 'CER-MDK-MPC' when referring to the linear Koopman baseline; please use the full name 'CER-MDK-MPC' consistently, since the CER module is part of that baseline controller.","section":"IV and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim depends on isolating the benefit of the bilinear structure, but the MPC implementation in Eq. (40) is a gain-scheduled linear model that also drops the curvature bilinear term. This is a technical gap that the authors can likely fix, but it requires re-evaluating the closed-loop results. The single-scenario HIL evaluation is thin for the strength of the claims made in the abstract. The reliance on the authors' own MDK-Net as a baseline is understandable but should be complemented by additional independent baselines and scenarios."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: a serious, incremental engineering paper. The genuinely new part is the deep bilinear Koopman model combined with a cumulative error regulator, applied to Frenet-frame vehicle control with driver inputs. The open-loop prediction results over 1,425 test trajectories are credible and show MDBK-Net beating the linear MDK and EDMDK baselines, especially in lateral states. The HIL setup on dSPACE with CarSim RT is real work, and the reported 0.58 ms per-step compute time is a solid real-time result.\n\nThe soft spots are real but not fatal. The biggest is in the MPC derivation. Equation (40) freezes the lifted state to its initial value Z_t for the whole horizon, which is the known quasi-LPV trick from Ref. [37]. But it also drops the curvature bilinear term H3(Z_t * kappa) entirely. That is not an approximation; it is an omission. Since curvature is an exogenous input and Z_t is known, H3(Z_t * kappa) is a known affine term that could be added for free. The result is that the controller's internal model is not even the local linearization of the learned bilinear model, and the paper's claim that it captures curvature nonlinearities in the control loop is unsupported. The improvement over CER-MDK still demonstrates the value of the H1 and H2 state-dependent input scheduling, so the bilinear benefit is not entirely an artifact, but the authors should include H3 or justify its absence.\n\nSecond, the closed-loop evaluation is a single double-lane-change maneuver at constant curvature, with no error bars or significance tests. 'Significant reductions' is too strong for a single run. The open-loop numbers are better supported because they average over many trajectories, but they also lack confidence intervals.\n\nThird, no code or data is released. For a data-driven method, that limits reproducibility, though it is not a correctness flaw.\n\nThe paper is clearly written and the authors engage honestly with prior literature. The self-citation to MDK-Net is fair because it is the direct predecessor. The bilinear matrices in Fig. 5 align with known physics, which is a nice check.\n\nBottom line: this is a useful incremental contribution for the Koopman-MPC community. It deserves serious peer review, not a desk reject. A referee should ask for the H3 term to be addressed, and for at least one more maneuver plus statistical reporting before acceptance.","headline":"Useful incremental engineering paper on deep bilinear Koopman MPC with a nice integral-error module; the closed-loop evidence is thin and the MPC derivation drops the curvature bilinear term.","tokens_in":19400,"tokens_out":5498,"would_cite":false,"duration_ms":58269,"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 learned bilinear Koopman model with cumulative error regulation enables a convex MPC that tracks paths up to 97 percent more accurately than EDMD and LTI baselines and solves each control step in 0.58 ms.","keywords":["Koopman operator","bilinear model","model predictive control","autonomous vehicles","Frenet frame","deep learning","cumulative error regulator","hardware-in-the-loop"],"falsifier":"Re-run the same CER-MDBK-MPC against the linear MDK baseline on a more aggressive maneuver, such as a higher-speed double lane change or a slalom with curvature near the training limits, and measure the deviation of the lifted state from its horizon-start value; if the bilinear controller's tracking advantage over the linear model shrinks or vanishes as that deviation grows, the convex linearization assumption is the limiting factor.","tokens_in":18314,"feed_emoji":"🚗","tokens_out":8164,"duration_ms":88334,"temperature":0.7,"pith_summary":"Koopman operator theory promises a way to treat nonlinear vehicle dynamics as linear dynamics in a lifted space, but standard lifts are hand-picked and tend to miss input–state coupling. This paper tries to remove that bottleneck by learning the lift and the dynamics together: a deep encoder produces observables, and a final Koopman layer learns a bilinear transition $Z_{k+1}=A Z_k+B U_k+\\sum_i H_i(U^i_k Z_k)$ that explicitly couples each control channel to the lifted state. Trained with a multi-step prediction loss, this model is embedded in a convex MPC by freezing the lifted state in the bilinear term over the prediction horizon, and augmented with a cumulative error regulator that propagates and penalizes accumulated tracking error. In hardware-in-the-loop tests against the paper's prior linear deep Koopman model, an Extended Dynamic Mode Decomposition (EDMD) Koopman model, and an LTI model, the controller reduces average tracking RMSE by 33–68 percent and keeps the per-step solve time at 0.58 ms, well inside the 25 ms control period. If the result holds, it shows that learned bilinear Koopman models are a practical route to accurate, convex, real-time vehicle control.","feed_headline":"Bilinear Koopman MPC cuts vehicle path error by up to 97 percent","feed_subtitle":"A learned bilinear Koopman model plus cumulative error control tracks curves accurately in 0.58 ms per step.","key_machinery":"The central object is the learned bilinear Koopman realization: an encoder $\\Phi_\\theta$ maps the six vehicle states to learnable observables, the lifted state is $Z_k=[x_k;\\Phi_\\theta(x_k)]$, and the transition is $Z_{k+1}=A Z_k+B U_k+\\sum_i H_i(U^i_k Z_k)$, where $U_k$ includes steering wheel angle, throttle/brake command, and road curvature. This structure makes input–state coupling a tunable part of the dynamics rather than a modeling afterthought. Around it, the Cumulative Error Regulator appends $e_{k+1}=e_k+(y_k-y_k^{\\mathrm{ref}})$ to the lifted state, and the MPC preserves convexity by replacing the lifted state inside the bilinear terms with its initial value $Z_t$ over the prediction horizon, producing the linearized update of Eq. 40.","core_discovery":"The central claim is that modeling the lifted vehicle dynamics as $Z_{k+1}=A Z_k+B U_k+\\sum_i H_i(U^i_k Z_k)$, where $Z_k$ concatenates the original Frenet-frame states with an encoder-learned observable vector and the matrices $H_i$ are learned end-to-end, captures input–state interactions that linear Koopman models miss. The paper supports this by showing that the learned $H_i$ matrices have sparsity patterns aligned with the nonlinear bicycle-model equations, by reporting lower open-loop prediction RMSE than the linear MDK-Net and EDMDK baselines, and by closing the loop with a convex MPC that freezes the lifted state in the bilinear terms to its horizon-start value and augments the state with a cumulative tracking-error integrator. On a double-lane-change maneuver with a high-fidelity vehicle simulator as the plant, the CER-MDBK-MPC cuts tracking error of path progress, lateral deviation, and heading error by 5–97 percent relative to EDMDK-MPC and by 35–77 percent relative to the linear MDK baseline, with average reductions of 33 percent, 38 percent, and 68 percent against the linear Koopman, EDMD, and LTI controllers.","pith_inferences":["A testable interpretation of the CER module is that it acts as output integral action on the tracking errors: at steady state the cumulative error must remain constant for the penalty to stay bounded, which cancels constant model-mismatch bias; the paper does not draw this equivalence explicitly.","Because the bilinear term is frozen at the horizon-start lifted state, the resulting controller is effectively a gain-scheduled LTV MPC whose input matrix changes with the current lifted state; comparing against an explicit LTV MPC built from Jacobian linearization would isolate what the learned bilinear schedule adds.","The same encoder-plus-bilinear-Koopman-layer-plus-error-integrator architecture should transfer to other actuated nonlinear systems such as manipulators or small aircraft, but that transfer is untested and the model's validity is bounded by the simulation distribution it was trained on."],"forward_implications":["The convex freezing trick turns the bilinear terms into a state-scheduled input matrix, so the same MPC formulation can be reused for any learned bilinear Koopman model without losing real-time solvability.","Open-loop, MDBK-Net reports 2-second prediction RMSE for lateral velocity and yaw rate reduced by 43 percent and 36 percent relative to MDK-Net, and by 66 percent and 50 percent relative to EDMDK, with the largest gains in states where input–state coupling matters.","Closed-loop, CER-MDBK-MPC reduces path progress, lateral deviation, and heading-error RMSE by 5 percent, 97 percent, and 88 percent versus EDMDK-MPC, and by 35 percent, 77 percent, and 59 percent versus the linear MDK baseline.","At 0.58 ms average solve time per step, the controller leaves roughly 24 ms of the sampling interval for sensing, communication, and safety checks, directly supporting the paper's real-time implementation claim.","The cumulative error regulator keeps pushing accumulated tracking errors toward zero over the whole horizon, which is why the largest visible improvements appear during and after the aggressive second lane change rather than only at steady state."],"supporting_citations":[{"why":"Supplies the operator-theoretic foundation: nonlinear dynamics become linear on an infinite-dimensional space of observables.","marker":"[14]"},{"why":"Provides the input-driven Koopman formulation and the linear-predictor MPC setting that the paper builds on.","marker":"[15]"},{"why":"Provides EDMD, the classical lifting baseline that the learned encoder is compared against.","marker":"[18]"},{"why":"Introduces bilinear Koopman realizations, the expressiveness-versus-tractability trade-off the paper adopts.","marker":"[22]"},{"why":"Earlier bilinear Koopman vehicle MPC whose structure and freezing trick are extended here to deep learned observables and Frenet-frame states.","marker":"[23]"},{"why":"Basis for learning Koopman observables with autoencoders instead of hand-picking a basis.","marker":"[28]"},{"why":"Supplies the multi-step prediction loss formulation that MDBK-Net adapts for long-horizon accuracy.","marker":"[30]"},{"why":"The prior deep Koopman Frenet-frame model (MDK-Net) that serves as the linear baseline and whose control structure is augmented with CER.","marker":"[31]"},{"why":"Provides the strategy of fixing the lifted state in bilinear terms over the prediction horizon to keep the MPC convex.","marker":"[37]"},{"why":"Provides the EDMDK-MPC baseline with polynomial liftings and road-curvature handling used for direct comparison.","marker":"[39]"}],"fun_headline_variants":["Bilinear Koopman MPC cuts vehicle path error by up to 97%","Learned bilinear Koopman lifting improves real-time vehicle MPC","Frenet-frame bilinear Koopman model reduces tracking error","Deep Koopman control for vehicles: up to 97% less error","Real-time bilinear Koopman MPC for precise vehicle tracking"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that holding the lifted state fixed at its starting value for the 20-step MPC horizon keeps the linearized bilinear model close to the true learned dynamics; if the vehicle maneuvers aggressively enough for the lifted state to leave that neighborhood, the model the controller optimizes no longer represents the bilinear dynamics, and the accuracy advantage would weaken—and the encoder is only as reliable as the simulation training distribution it was learned from.","fun_headline_variants_meta":{"raw":{"variants":["Bilinear Koopman MPC cuts vehicle path error by up to 97%","Learned bilinear Koopman lifting improves real-time vehicle MPC","Frenet-frame bilinear Koopman model reduces tracking error","Deep Koopman control for vehicles: up to 97% less error","Real-time bilinear Koopman MPC for precise vehicle tracking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000729,"raw_usage":{"total_tokens":3313,"prompt_tokens":1044,"completion_tokens":2269,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":2174}},"tokens_in":660,"tokens_out":2269,"duration_ms":18530,"temperature":1.0,"reasoning_tokens":2174,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:44:16.492108+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same CER-MDBK-MPC against the linear MDK baseline on a more aggressive maneuver, such as a higher-speed double lane change or a slalom with curvature near the training limits, and measure the deviation of the lifted state from its horizon-start value; if the bilinear controller's tracking advantage over the linear model shrinks or vanishes as that deviation grows, the convex linearization assumption is the limiting factor.","supporting_citations":[{"cited_title":"Hamiltonian systems and transformation in hilbert space,","cited_arxiv_id":null,"evidence_quote":"Supplies the operator-theoretic foundation: nonlinear dynamics become linear on an infinite-dimensional space of observables."},{"cited_title":"Linear predictors for nonlinear dynamical sys- tems: Koopman operator meets model predictive control,","cited_arxiv_id":null,"evidence_quote":"Provides the input-driven Koopman formulation and the linear-predictor MPC setting that the paper builds on."},{"cited_title":"Extending data-driven koopman analysis to actuated systems,","cited_arxiv_id":null,"evidence_quote":"Provides EDMD, the classical lifting baseline that the learned encoder is compared against."},{"cited_title":"Autonomous driving using linear model predictive control with a koopman operator based bilinear vehicle model,","cited_arxiv_id":null,"evidence_quote":"Earlier bilinear Koopman vehicle MPC whose structure and freezing trick are extended here to deep learned observables and Frenet-frame states."},{"cited_title":"Deep learning for universal linear embeddings of nonlinear dynamics,","cited_arxiv_id":null,"evidence_quote":"Basis for learning Koopman observables with autoencoders instead of hand-picking a basis."},{"cited_title":"Ddk: A deep koopman approach for longitudinal and lateral control of autonomous ground vehicles,","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-step prediction loss formulation that MDBK-Net adapts for long-horizon accuracy."},{"cited_title":"Multi- step deep koopman network (mdk-net) for vehicle control in frenet frame,","cited_arxiv_id":null,"evidence_quote":"The prior deep Koopman Frenet-frame model (MDK-Net) that serves as the linear baseline and whose control structure is augmented with CER."},{"cited_title":"Deep bilinear koopman model predictive control for nonlinear dynamical systems,","cited_arxiv_id":null,"evidence_quote":"Provides the strategy of fixing the lifted state in bilinear terms over the prediction horizon to keep the MPC convex."},{"cited_title":"Koopman operator-based model identification and control for automated driving vehicle,","cited_arxiv_id":null,"evidence_quote":"Provides the EDMDK-MPC baseline with polynomial liftings and road-curvature handling used for direct comparison."}],"review_version":1}