Pith. sign in

REVIEW 4 major objections 5 minor 30 references

Learning Based MPC for Autonomous Driving Using a Low Dimensional Residual Model

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper establishes that the residual between a single-track vehicle model and the real vehicle can be learned by a Gaussian process using only front slip angle, rear slip angle, and command torque as features, and that embedding this…

desk verdict A reasonable engineering paper whose central 'low-dimensional' claim is under-tested: no 5-feature baseline, no check for omitted vx-dependence in the residual. read the letter →

arxiv 2412.03874 v1 pith:3HQAPP65 submitted 2024-12-05 cs.RO cs.SYeess.SY

Yaoyu Li , Chaosheng Huang , Dongsheng Yang , Wenbo Liu , Jun Li This is my paper · ORCID
classification cs.ROcs.SYeess.SY
keywords learning-basedMPCGaussianprocessresidualmodelvehicledynamicsautonomousracingfeaturedimensionalityreductionpredictivecontouringcontrolsingle-tracktireslipangle
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that the mismatch between a single-track vehicle model and the real car can be captured by a low-dimensional Gaussian process using just three quantities: front slip angle $\alpha_f$, rear slip angle $\alpha_r$, and command torque $T$. This is a sharp reduction from the five dynamic states used in earlier residual-model work, and it makes learning-based MPC tractable on real vehicles. The authors embed this residual model in a model predictive contouring controller and validate it in simulation and on a test vehicle, reporting faster lap times and substantially smaller prediction errors on lateral velocity and yaw rate.

What carries the argument

The load-bearing object is the low-dimensional residual model $g(z)$ with feature vector $z = [\alpha_f,\alpha_r,T]^T$, implemented as independent Gaussian processes for the three velocity states. It is connected to the nominal model through the discrete-time update $x_{k+1} = f(x_k,u_k) + B_d g(z_k)$, where $B_d$ selects the $v_x$, $v_y$, and $\omega$ subspace. The dimensionality reduction is justified by the small-angle approximation in Equation (9), which leaves the velocity-state errors as sums of front and rear tire-force deviations; the claim that those deviations depend only on the three selected features is what makes the GP tractable. The feature-space constraints — the friction ellipse of Equation (16), the slip-angle bounds of Equation (17), and the $\alpha_f-\alpha_r$ handling-diagram bound of Equation (18) — explicitly define the valid region in feature space and shrink the set of operating points the training data must cover.

What would settle it

Run the same vehicle at two very different speeds (for example 10 m/s and 25 m/s) through a corner that produces identical values of $(\alpha_f,\alpha_r,T)$. If the residuals in $v_x$, $v_y$, and $\omega$ measured at these matched feature points differ by more than the GP's predicted uncertainty, the three-feature assumption is falsified, because speed-dependent effects such as load transfer and downforce are not represented in the feature vector.

Watch

Extended reading notes

Core claim

The paper's central claim is that the deviations of the nominal single-track model on the velocity states $v_x$, $v_y$, and $\omega$ can be expressed as a function of only the front slip angle $\alpha_f$, the rear slip angle $\alpha_r$, and the command torque $T$. This follows from decomposing the nominal model into invariable elements (mass, inertia, geometry, drag) that are calibrated experimentally, and variable elements (tire forces) whose errors are learned. With the small-angle approximation in Equation (9), the velocity-state residuals simplify to combinations of tire-force deviations, and since lateral force deviations depend primarily on slip angles while longitudinal force deviations depend primarily on torque, the three features suffice. The residual model $g(z)$ is trained as independent Gaussian processes on these features and attached through $x_{k+1} = f(x_k,u_k) + B_d g(z_k)$; feature-space constraints derived from the friction ellipse, slip-angle limits, and the vehicle handling diagram define a valid operating region. In experiments, adding this residual model cuts mean yaw-rate prediction error from about $1.88\times10^{-2}$ rad/s to $0.87\times10^{-2}$ rad/s on the real vehicle, and reduces lap time from 136.38 s to 131.40 s.

Load-bearing premise

The whole benefit rests on the assumption that tire-force errors, and therefore the residual in the velocity states, depend only on front and rear slip angle and commanded torque, so no other state such as speed, yaw rate, or steering angle needs to be included in the learned model.

Editorial extensions

If this is right

  • With only three features, the training set required to cover the feature space is small enough to be collected in a single lap, making GP-based residual MPC practical on real vehicles.
  • The feature-space constraints keep the learned model inside a physically valid region, so the controller will not extrapolate to unsafe slip-angle or tire-force combinations.
  • The residual model adapts continuously as data accumulate each lap; in simulation the lap time improves from 119.63 s to 104.85 s over five laps of updates.
  • Model prediction errors on $v_y$ and $\omega$ drop substantially after the GP residual is added; in the hardware experiment the mean yaw-rate error falls from $1.88\times10^{-2}$ rad/s to $0.87\times10^{-2}$ rad/s.
  • The framework provides a concrete way to combine physics-based nominal models with data-driven compensation, reducing the burden of collecting exhaustive training data for high-dimensional feature spaces.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same three-feature argument could extend to other vehicle models whose residuals enter the velocity states through tire-force deviations, provided the small-angle approximation holds; a useful test is to check whether residuals at matched $(\alpha_f,\alpha_r,T)$ points still spread with $v_x$ or $v_y$.
  • The handling-diagram constraint on $\alpha_f-\alpha_r$ may double as a safety envelope in mixed-friction conditions, since it implicitly bounds the steady-state lateral acceleration the model can represent.
  • A direct extension of this work would be to replace the fixed kernel hyperparameters with online adaptation, allowing the GP to track slow changes in tire wear or road surface without rebuilding the training set.
  • If the features truly capture the dominant error modes, the same low-dimensional residual could be transferred across vehicle platforms after only a short calibration drive, because the slip-angle and torque features are physically normalized.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a learning-based Model Predictive Contouring Controller (MPCC) for autonomous racing in which the nominal single-track vehicle model is augmented by a Gaussian Process (GP) residual model. The main claimed contribution is that the residual can be represented with only three features—front slip angle, rear slip angle, and commanded torque—instead of the five dynamic states (vx, vy, omega, delta, T) used in earlier GP-residual work. Physical constraints among the features are formulated to define a valid region of the feature space. The method is evaluated in CarSim simulation and on a real vehicle during two laps; the authors report consistent reductions in one-step prediction error and lap time relative to a nominal-model-only MPC.

Significance. If the three-feature sufficiency claim were established, the method would make GP-based residual MPC considerably more tractable for real-time autonomous racing, reducing both data coverage requirements and online computational cost. The evaluation protocol is not circular: the GP is trained on data from earlier laps and evaluated on later laps. The empirical results are substantial and consistent — Tables I–IV show meaningful reductions in lap time and in v_y and omega prediction errors in both simulation and hardware. However, the central claim that a low-dimensional feature set captures the residual is not directly tested: no comparison with the previous five-feature model is reported, and the physical derivation in Section II.C does not exclude omitted-state dependence. The paper is therefore promising and within scope, but the load-bearing claim needs additional evidence.

major comments (4)
  1. [Section II.C, Eqs. (8)–(9) and (15)] The derivation does not establish the central premise that the tire-force deviations depend only on alpha_f, alpha_r, and T. Equation (9) is only a small-angle linearization that expresses the velocity residuals as linear combinations of the force deviations; it says nothing about which physical variables determine Delta F_f/r,x/y. In the very model defined by the authors, lateral force depends on normal load, which changes with longitudinal/lateral load transfer and, through the Cw vx^2 drag term in Eq. (6), with speed; longitudinal force depends on wheel-slip dynamics that are not determined by commanded torque T alone but also by vx. Consequently, for a fixed triple (alpha_f, alpha_r, T), the state vx can vary and the residual can vary with it, so the GP is a regression with an omitted variable. The paper should either provide a theoretical argument that these dependencies are negligible over the tested operating regime, or empirically show residual constancy within feature-space bins across varying vx. This is load-bearing because the claimed advantage over the five-feature model rests entirely on this reduction.
  2. [Section IV, Tables I–IV] The paper never compares the proposed three-feature residual model with the five-feature model [vx, vy, omega, delta, T] that it claims to supersede. Tables II and IV compare the nominal model only with the three-feature GP residual model. Without a five-feature baseline trained on the same data, the statement that the feature dimensionality is reduced 'from 5 to 3' is a claim about complexity, not a measured result about accuracy. A direct comparison of prediction error and controller performance between the three-feature and five-feature GPs should be added; if the five-feature model gives comparable or better results, the low-dimensional advantage is not demonstrated.
  3. [Section III.B, Eq. (20) and Section IV.A] The MPC formulation appears to use only the posterior mean of the GP. Equation (14) defines the posterior distribution N(mu, Sigma), but Eq. (20) includes only Bdg(zk) in the dynamics and no term involving Sigma_a(zk) in the cost or constraints. If the controller ignores the posterior variance, this should be stated explicitly and the relevance of the probabilistic formalism to the controller should be clarified; if the variance is used through chance constraints or a robustness penalty, the implementation must be described. As written, the reader cannot tell what role GP uncertainty plays, which is also relevant to the claim that the feature constraints keep the controller inside the learned valid region.
  4. [Section II.D, Eq. (16)–(18) and Section IV.B/C] The feature constraints are used both to discard training data outside the valid region and as soft constraints in Eq. (20). Because the constraints are soft, the optimizer may select states outside the valid region, but the GP is never trained on those states, so the residual model is then queried in extrapolation. The interaction between soft constraints and the discarded training data is not analyzed, and no safeguard (e.g., variance-based caution or constraint tightening) is described. This matters because the entire low-dimensional argument depends on the model being used inside a carefully restricted region.
minor comments (5)
  1. [Eq. (20)] The track-radius constraint is printed as ||X_k − X_c(theta_k)||^2 + ||X_k − X_c(theta_k)||^2 <= R^2; the second norm should presumably involve Y_k. This typo obscures the actual constraint.
  2. [Section IV.A] The GP hyperparameters (M, sigma_f^2, sigma_w) and the sparsification threshold for the training set are not reported; these are needed for reproducibility.
  3. [Section IV.A] The achieved computation time per MPC iteration is not reported, even though real-time feasibility is stated as a motivation; a short table or paragraph on solver runtime with and without the residual model would be useful.
  4. [Fig. 6(a)] The label 'GG-diagram' is nonstandard; the conventional term is 'g-g diagram'.
  5. [Section IV.C, Conclusion] The statement that a training set covering only the constrained valid region delivers 'robust performance across varied operating conditions' is stronger than the evidence: the experiments use a single track, a single surface, and a limited speed range, with no cross-track or cross-condition generalization test.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the GP residual model is trained on data from earlier laps and evaluated on subsequent laps, and the three-feature selection is a physical modeling assumption rather than a re-statement of the target result.

full rationale

The derivation chain is self-contained. The residual labels y_k are computed from measured state transitions minus the nominal model (Eq. 10), and the GP is trained on lap 1 data and evaluated on later laps (simulation Table II, hardware Table IV), so the reported error reductions are genuine out-of-sample predictions rather than re-fitted values. The reduction from five features (vx, vy, omega, delta, T) to three features (alpha_f, alpha_r, T) is justified by a physical argument in Section II.C: Eq. (9) expresses velocity residuals as linear combinations of tire-force deviations, and the paper asserts that tire-force deviations are primarily functions of slip angles and command torque. Whether this assertion is fully accurate (e.g., load transfer, speed-dependent downforce, or wheel-slip dynamics) is a model-adequacy and generalization concern, not a circularity: the paper does not define the residual as a function of these three features by construction, nor does it fit the features to the reported performance metric. The citations to prior work ([26], [28], [30]) support algorithmic components (sparsification, MPCC formulation, QP solver) and are not load-bearing self-citations. No equation or claim reduces the predicted result to its inputs, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a small number of domain assumptions about which variables govern tire-force errors. No new physical entities are introduced. Several free parameters (GP hyperparameters, constraint limits, MPC weights) are not specified, which affects reproducibility but not the logical validity of the method.

free parameters (4)
  • GP hyperparameters (M, sigma_f^2, sigma_w) = not specified
    The SE kernel length-scales M, signal variance sigma_f^2, and observation noise sigma_w are learned from training data, but no values or fitting procedure are given.
  • Feature constraint parameters (p_long, p_ellipse, alpha_max, Delta_alpha_max) = not specified
    These define the valid region in feature space and filter training and control data; values are chosen by hand and not reported.
  • MPC weights (q_l, q_c, q_v, R_x, R_u) = not specified
    Cost weights and regularization matrices determine controller behavior; values are not provided and appear hand-tuned.
  • Training set size m and sparsification threshold = 100 points, threshold unspecified
    The data selection mechanism from [26] is applied, but the independence threshold is not given; m=100 is stated.
assumptions (5)
  • domain assumption Model errors affect only the velocity states vx, vy, omega; kinematic states X,Y,phi and actuator states T,delta are accurate.
    Stated in Section II.B: 'the dynamics of these states are assumed to be sufficiently accurate, with model errors only affecting the velocity states.' This justifies the structure of Bd and the residual model output.
  • domain assumption Tire force deviations depend only on alpha_f, alpha_r, and T, with coupling effects captured by these variables.
    Section II.C argues from Eq (9) that delta F_f/r,y depends on slip angles and delta F_f/r,x on torque, then selects z=[alpha_f,alpha_r,T]. This neglects explicit dependence on vx, vy, omega, delta, and load transfer.
  • domain assumption Small steering angle approximation: cos(delta) is approximately 1 and higher-order terms are negligible.
    Used in Eq (9) to simplify model error; may be violated during high-lateral-acceleration racing where delta is not very small.
  • domain assumption Gaussian process residuals are independent across output dimensions and use a squared exponential kernel.
    Section II.C states 'each dimension of the residual model output is assumed to be uncorrelated with the others.' This simplifies the GP but may not hold for coupled tire forces.
  • ad hoc to paper Data outside the feature constraints are invalid and can be excluded from training and control.
    Section II.D states data violating constraints are 'considered invalid and excluded'; this is a modeling choice that could bias the residual model if the vehicle operates near or beyond those limits.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Based MPC for Autonomous Driving Using a Low Dimensional Residual Model." pith.science (2026). https://pith.science/paper/3HQAPP65

@misc{pith2026241203874,
  author       = {Pith},
  title        = {Pith review of: Learning Based MPC for Autonomous Driving Using a Low Dimensional Residual Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3HQAPP65}},
  note         = {Machine review of arXiv:2412.03874}
}
read the original abstract

In this paper, a learning based Model Predictive Control (MPC) using a low dimensional residual model is proposed for autonomous driving. One of the critical challenge in autonomous driving is the complexity of vehicle dynamics, which impedes the formulation of accurate vehicle model. Inaccurate vehicle model can significantly impact the performance of MPC controller. To address this issue, this paper decomposes the nominal vehicle model into invariable and variable elements. The accuracy of invariable component is ensured by calibration, while the deviations in the variable elements are learned by a low-dimensional residual model. The features of residual model are selected as the physical variables most correlated with nominal model errors. Physical constraints among these features are formulated to explicitly define the valid region within the feature space. The formulated model and constraints are incorporated into the MPC framework and validated through both simulation and real vehicle experiments. The results indicate that the proposed method significantly enhances the model accuracy and controller performance.

Figures

Figures reproduced from arXiv: 2412.03874 by the authors.

Figure 1
Figure 1. The learning based MPC framework. vehicle dynamics can vary during operation due to factors such as temperature fluctuations and tire wear, rendering pre-defined physical models insufficiently adaptive to these dynamics changes [11]. In recent years, data-driven algorithms have received sig￾nificant attention in the academic community due to their excellent ability to approximate unknown functions [12], [13]. These … view at source ↗
Figure 2
Figure 2. Vehicle dynamics model. essential vehicle dynamics, as formulated below: x˙ =             vxcosφ − vysinφ vxsinφ + vycosφ ω 1 m (Fr,x − Fd − Ff,ysinδ + Ff,xcosδ) + vyω 1 m (Fr,y + Ff,ycosδ + Ff,xsinδ) − vxω 1 Iz ((Ff,ycosδ + Ff,xsinδ)lf − Fr,ylr) ∆T ∆δ             . (2) Here, Ff,x and Fr,x denote the longitudinal force on the front and rear wheels, respectively, while Ff,y and Fr,y denote the… view at source ↗
Figure 3
Figure 3. Handling diagram of vehicle. functions of the tire side slip angles; therefore, deviations in lateral tire forces, ∆Ff/r,y, are primarily functions of the tire side slip angle αf/r. Similarly, deviations in longitudinal tire force, ∆Ff/r,x, are primarily due to differences between the tire force predicted based on the commanded torque and actual tire force, making ∆Ff/r,x primarily a function of the command torque T… view at source ↗
Figures from the paper (6 more)
Figure 6
Figure 6. Figure 6: Comparison of nominal (lap 1) and learning based [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 5
Figure 5. Figure 5: Racing trajectory of lap 1 and lap 6. TABLE I: Experimental Results in Simulation Lap Time (s) ∥ay∥max (g) Data Updates Average Speed (m/s) 1 119.63 0.59 - 19.46 2 111.20 0.68 82 20.94 3 107.60 0.74 78 21.64 4 105.43 0.71 20 22.08 5 104.93 0.72 6 22.18 6 104.85 0.73 9 …
Figure 8
Figure 8. Figure 8: Comparison of nominal (lap 1) and learning based MPC controller (lap 2). [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 9
Figure 9. Figure 9: Model error and GP prediction in lap 2 TABLE IV: Model Error in Hardware Experiment Lap evy (10−2m/s) eω (10−2 rad/s) nominal nominal nominal nominal model +residual model +residual 1 4.21±3.18 - 1.88±1.33 - 2 4.40±3.08 1.57±1.46 1.88±1.39 0.87±0.72 −0.1 0.0 0.1 αf[rad…
Figure 10
Figure 10. Figure 10: Data distribution within the feature space [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]
Figure 11
Figure 11. Figure 11: The real vehicle testing platform [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 23 canonical work pages

  1. [1]

    Towards the unified principles for level 5 autonomous vehicles,

    J. Wang, H. Huang, K. Li, and J. Li, “Towards the unified principles for level 5 autonomous vehicles,” Engineering, vol. 7, no. 9, pp. 1313– 1325, 2021

    work page 2021

  2. [2]

    Model predictive path tracking control for automated road vehicles: A review,

    P. Stano, U. Montanaro, D. Tavernini, M. Tufo, G. Fiengo, L. Novella, and A. Sorniotti, “Model predictive path tracking control for automated road vehicles: A review,” Annual reviews in control , vol. 55, pp. 194– 236, 2023

    work page 2023

  3. [3]

    Implementation of mpc-based path tracking for autonomous vehicles considering three vehicle dynamics models with different fidelities,

    S. Chen, H. Chen, and D. Negrut, “Implementation of mpc-based path tracking for autonomous vehicles considering three vehicle dynamics models with different fidelities,” Automotive Innovation, vol. 3, no. 4, pp. 386–399, 2020

    work page 2020

  4. [4]

    An overview on vehicle dynamics,

    S. Yang, Y . Lu, and S. Li, “An overview on vehicle dynamics,” International Journal of Dynamics and Control , vol. 1, pp. 385–395, 2013

    work page 2013

  5. [5]

    Vehicle dynamics,

    D. Schramm, M. Hiller, and R. Bardini, “Vehicle dynamics,” Modeling and Simulation. Berlin, Heidelberg , vol. 151, 2014

    work page 2014

  6. [6]

    Feedback control of a nonholonomic car-like robot,

    A. De Luca, G. Oriolo, and C. Samson, “Feedback control of a nonholonomic car-like robot,” Robot motion planning and control , pp. 171–253, 2005

    work page 2005

  7. [7]

    Robust lateral trajectory following control of unmanned vehicle based on model predictive control,

    H. Gao, Z. Kan, and K. Li, “Robust lateral trajectory following control of unmanned vehicle based on model predictive control,” IEEE/ASME Transactions on Mechatronics , vol. 27, no. 3, pp. 1278–1287, 2021

    work page 2021

  8. [8]

    Output constraint control on path following of four-wheel independently actuated autonomous ground vehicles,

    C. Hu, R. Wang, F. Yan, and N. Chen, “Output constraint control on path following of four-wheel independently actuated autonomous ground vehicles,” IEEE Transactions on V ehicular Technology, vol. 65, no. 6, pp. 4033–4043, 2015

    work page 2015

Show all 30 references
  1. [9]

    Coordinating tire forces to avoid obsta- cles using nonlinear model predictive control,

    M. Brown and J. C. Gerdes, “Coordinating tire forces to avoid obsta- cles using nonlinear model predictive control,” IEEE Transactions on Intelligent V ehicles, vol. 5, no. 1, pp. 21–31, 2019

    2019

  2. [10]

    Time- optimal trajectory planning for a race car considering variable tyre- road friction coefficients,

    F. Christ, A. Wischnewski, A. Heilmeier, and B. Lohmann, “Time- optimal trajectory planning for a race car considering variable tyre- road friction coefficients,” V ehicle system dynamics, vol. 59, no. 4, pp. 588–612, 2021

    2021

  3. [11]

    Neural network model predictive motion control applied to automated driving with unknown friction,

    N. A. Spielberg, M. Brown, and J. C. Gerdes, “Neural network model predictive motion control applied to automated driving with unknown friction,” IEEE Transactions on Control Systems Technology , vol. 30, no. 5, pp. 1934–1945, 2021

    1934

  4. [12]

    Robots that can adapt like animals,

    A. Cully, J. Clune, D. Tarapore, and J.-B. Mouret, “Robots that can adapt like animals,” Nature, vol. 521, no. 7553, pp. 503–507, 2015

    2015

  5. [13]

    Learning-based model predictive control: Toward safe learning in control,

    L. Hewing, K. P. Wabersich, M. Menner, and M. N. Zeilinger, “Learning-based model predictive control: Toward safe learning in control,” Annual Review of Control, Robotics, and Autonomous Sys- tems, vol. 3, no. 1, pp. 269–296, 2020

    2020

  6. [14]

    A survey of deep learning applications to autonomous vehicle control,

    S. Kuutti, R. Bowden, Y . Jin, P. Barber, and S. Fallah, “A survey of deep learning applications to autonomous vehicle control,” IEEE Transactions on Intelligent Transportation Systems , vol. 22, no. 2, pp. 712–733, 2020

    2020

  7. [15]

    Neural network vehicle models for high-performance automated driving,

    N. A. Spielberg, M. Brown, N. R. Kapania, J. C. Kegelman, and J. C. Gerdes, “Neural network vehicle models for high-performance automated driving,” Science robotics , vol. 4, no. 28, p. eaaw1975, 2019

    2019

  8. [16]

    Fusion of machine learning and mpc under uncertainty: What advances are on the horizon?

    A. Mesbah, K. P. Wabersich, A. P. Schoellig, M. N. Zeilinger, S. Lucia, T. A. Badgwell, and J. A. Paulson, “Fusion of machine learning and mpc under uncertainty: What advances are on the horizon?” in 2022 American Control Conference (ACC) . IEEE, 2022, pp. 342–357

    2022

  9. [17]

    Scalable deep kernel gaussian process for vehicle dynamics in autonomous racing,

    J. Ning and M. Behl, “Scalable deep kernel gaussian process for vehicle dynamics in autonomous racing,” in Conference on Robot Learning. PMLR, 2023, pp. 761–773

    2023

  10. [18]

    Learning-based nonlinear model predictive control to improve vision-based mobile robot path-tracking in challenging outdoor environments,

    C. J. Ostafew, A. P. Schoellig, and T. D. Barfoot, “Learning-based nonlinear model predictive control to improve vision-based mobile robot path-tracking in challenging outdoor environments,” in 2014 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2014, ...

    2014

  11. [19]

    Robust constrained learning-based nmpc enabling reliable mobile robot path tracking,

    C. J. Ostafew, A. P. Schoellig, and T. D. Barfoot, “Robust constrained learning-based nmpc enabling reliable mobile robot path tracking,” The International Journal of Robotics Research , vol. 35, no. 13, pp. 1547–1563, 2016

    2016

  12. [20]

    Learning multimodal models for robot dynamics online with a mixture of gaussian process experts,

    C. D. McKinnon and A. P. Schoellig, “Learning multimodal models for robot dynamics online with a mixture of gaussian process experts,” in 2017 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2017, pp. 322–328

    2017

  13. [21]

    Cautious nmpc with gaussian process dynamics for autonomous miniature race cars,

    L. Hewing, A. Liniger, and M. N. Zeilinger, “Cautious nmpc with gaussian process dynamics for autonomous miniature race cars,” in 2018 European Control Conference (ECC) . IEEE, 2018, pp. 1341– 1348

    2018

  14. [22]

    Cautious model predictive control using gaussian process regression,

    L. Hewing, J. Kabzan, and M. N. Zeilinger, “Cautious model predictive control using gaussian process regression,” IEEE Transactions on Control Systems Technology , vol. 28, no. 6, pp. 2736–2743, 2019

    2019

  15. [23]

    Amz driverless: The full autonomous racing system,

    J. Kabzan, M. I. Valls, V . J. Reijgwart, H. F. Hendrikx, C. Ehmke, M. Prajapat, A. B ¨uhler, N. Gosala, M. Gupta, R. Sivanesan et al. , “Amz driverless: The full autonomous racing system,” Journal of Field Robotics, vol. 37, no. 7, pp. 1267–1294, 2020

    2020

  16. [24]

    C. K. Williams and C. E. Rasmussen, Gaussian processes for machine learning. MIT press Cambridge, MA, 2006, vol. 2, no. 3

    2006

  17. [25]

    Learning- based model predictive control for autonomous racing,

    J. Kabzan, L. Hewing, A. Liniger, and M. N. Zeilinger, “Learning- based model predictive control for autonomous racing,” IEEE Robotics and Automation Letters , vol. 4, no. 4, pp. 3363–3370, 2019

    2019

  18. [26]

    Incremental online sparsification for model learning in real-time robot control,

    D. Nguyen-Tuong and J. Peters, “Incremental online sparsification for model learning in real-time robot control,” Neurocomputing, vol. 74, no. 11, pp. 1859–1867, 2011

    2011

  19. [27]

    Bifurcation analysis of an automobile model negotiating a curve,

    F. D. Rossa, G. Mastinu, and C. Piccardi, “Bifurcation analysis of an automobile model negotiating a curve,” V ehicle system dynamics, vol. 50, no. 10, pp. 1539–1562, 2012

    2012

  20. [28]

    Optimization-based au- tonomous racing of 1: 43 scale rc cars,

    A. Liniger, A. Domahidi, and M. Morari, “Optimization-based au- tonomous racing of 1: 43 scale rc cars,” Optimal Control Applications and Methods , vol. 36, no. 5, pp. 628–647, 2015

    2015

  21. [29]

    Pacejka, Tire and vehicle dynamics

    H. Pacejka, Tire and vehicle dynamics . Elsevier, 2005

    2005

  22. [30]

    Hpipm: a high-performance quadratic programming framework for model predictive control,

    G. Frison and M. Diehl, “Hpipm: a high-performance quadratic programming framework for model predictive control,” IF AC- PapersOnLine, vol. 53, no. 2, pp. 6563–6569, 2020

    2020

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.