A Contingency Model Predictive Control Framework for Safe Learning

Alexander Katriniok; Maurice Heemels; Merlijne Geurts; Tren Baltussen

REVIEW 2 major objections 3 minor 21 references

A Contingency Model Predictive Control Framework for Safe Learning

T0 review · 2 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper introduces a two-horizon contingency MPC scheme that couples a worst-case safety plan with a data-driven performance plan, and proves the combined controller inherits the safety plan's recursive feasibility.

desk verdict Useful CMPC framework, but the hard-constrained inheritance theorem has a real gap; the soft-constrained variant and simulations hold up. read the letter →

arxiv 2505.22776 v1 pith:JP5OUPOA submitted 2025-05-28 math.OC

Merlijne Geurts , Tren Baltussen , Alexander Katriniok , Maurice Heemels This is my paper

classification math.OC MSC 93B4593C1093C55

keywords contingencymodelpredictivecontrolsafelearningrobustrecursivefeasibilitylearning-basedMPCGaussianprocessconstraintsatisfactionautomatedlanemerginguncertainnonlinearsystems

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

This paper proposes a general contingency model predictive control (CMPC) framework in which a robust MPC, responsible for worst-case safety, and a learning-based MPC, responsible for performance, run as two coupled horizons sharing the same current state and first input. The central claim is that, under two stated assumptions, the combined optimal control problem is robustly recursively feasible: if it is solvable at time $k$, it remains solvable at time $k+1$ for every allowed disturbance, so the safety guarantee of the robust controller carries over to the learning-based controller. The authors prove this inheritance for hard constraints and, when the learned model is unreliable, for a soft-constrained variant that uses slack variables. The framework is instantiated by combining a robust lane-merging controller with a Gaussian-process MPC, and the simulation shows the combined controller learning to merge more assertively than standalone robust MPC while preserving constraint satisfaction.

What carries the argument

The central object is the two-horizon optimal control problem (2), which solves for a robust input sequence $\bar{U}_k$ and a learning-based input sequence $\hat{U}_k$ under two coupled prediction models: a nominal model (2b) for the contingency horizon and a learned-augmented model (2c) for the performance horizon. The horizons are coupled only through the shared initial state constraint (2d) and the shared first input (2e). The mechanism that carries the safety argument is trajectory reuse: robust recursive feasibility of the RMPC (Assumption 1) ensures a shifted robust solution exists at the next time step, and Assumption 2 is intended to let the performance horizon copy that shifted solution. The tightened constraint sets $\bar{\mathbb{X}}_j$ and the performance sets $\hat{\mathbb{X}}_{j|k}$ encode the safety margin, while the learned term $d(\cdot,\mathcal{D}_k)$ in (2c) is what carries the performance improvement.

What would settle it

A direct test is to start from a robustly feasible state, take the shifted robust input sequence, and compute the performance model's predicted states with the learned term $d$; if any predicted state leaves its performance set $\hat{\mathbb{X}}_{j|k}$ while each robust nominal state lies in $\bar{\mathbb{X}}_j$ and $d \in \mathcal{W}$, the reuse construction in Theorem 1 fails. Concretely, choose $\hat{\mathbb{X}}_{j|k} = \bar{\mathbb{X}}_j$ (allowed by Assumption 2) and a nonzero $d$ in $\mathcal{W}$, then check the performance prediction after one step from a robust nominal state at the boundary of $\bar{\mathbb{X}}_1$.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 1: if the two-horizon optimal control problem (2) is feasible at time $k$, then under Assumptions 1 and 2 it remains feasible at time $k+1$ for every disturbance $w_k \in \mathcal{W}$. The proof constructs the future solution by taking the robust MPC's shifted nominal trajectory and using it for both the contingency and the performance horizon. Theorem 2 extends the guarantee to the soft-constrained problem (5), where slack variables absorb violations of the performance-horizon constraints and only Assumption 1 is needed. The lane-merging case study then shows the intended consequence: the combined controller keeps the robust safety guarantee while using a Gaussian-process estimate of the other driver's acceleration to merge more assertively and at lower average cost than the standalone robust controller.

Load-bearing premise

The hard-constrained theorem depends on the performance horizon's admissible sets being robust enough that replaying the safety plan with the learned model's extra term still stays inside them; the stated containment condition alone does not guarantee this.

Editorial extensions

If this is right

Any robust MPC satisfying Assumption 1 can be paired with any learning-based MPC whose learned residual lies in $\mathcal{W}$ and whose constraint sets contain the robust sets, and the hybrid inherits the robust MPC's safety guarantee without redesigning the learner.
The soft-constrained variant of Theorem 2 keeps recursive feasibility even when the learned model is wrong, so online data collection and model updates cannot destroy the safety property.
Earlier contingency MPC schemes, including the robust-constraint learning-based MPC of [7] and the invariant-set contingency MPC of [15], become special cases of one framework with a common feasibility proof.
In the lane-merging study, CMPC merges in front in 88 of 231 scenarios versus 72 for standalone RMPC, lowers the average closed-loop cost from 2.89 to 2.14, and keeps slack at $10^{-7}$ to $10^{-6}$, whereas standalone GP-MPC violates the safety constraint by up to 0.98 m.
The framework can flag initial conditions for which no robust control invariant terminal set is reachable, so safety cannot be guaranteed and the scenario should be handled before deployment.

Reading between the lines

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

The proof of Theorem 1 silently requires a stronger containment than Assumption 2 states: the performance constraint sets must be robust to the learned disturbance along the reused trajectory, essentially $\bar{\mathbb{X}}_j \oplus \mathcal{W} \subseteq \hat{\mathbb{X}}_{j|k}$; otherwise a performance prediction can leave its allowed set even when the robust plan is feasible.
The framework suggests a modular recipe for safe learning: keep the worst-case safety plan fixed and wrap any learned model whose residual is bounded by $\mathcal{W}$, so safety certification does not require analyzing the learner's internal behavior.
A practical extension is to read the slack $\|E_k\|_1$ in (5) as an online confidence signal: persistent positive slack indicates the learned model has left its validity region, and the blending weight $P$ in the cost can be shifted toward the contingency horizon.
The same coupling idea could be applied with a stochastic performance horizon, such as chance-constrained GP-MPC, while the contingency horizon remains deterministic, yielding a safety filter that does not interrupt the learning loop.

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.

Desk Editor's Note

read the letter

Here's my read on arXiv:2505.22776. The paper's contribution is a conceptual multi-horizon contingency MPC framework that couples a robust MPC with a learning-based MPC, and it claims an inheritance theorem: if the RMPC is robustly recursively feasible, then so is the CMPC. That theorem is the centerpiece, and it has a real gap.

What's genuinely new: the abstract coupling of classes of RMPC and LB-MPC in a single two-horizon OCP with shared first input. Prior CMPC works were more specific. The paper also does a good job of laying out the framework, positioning it relative to [11], [13]-[15], and giving a concrete implementation with GP-MPC and a lane-merging simulation. The soft-constrained variant (Theorem 2) is sound, and the simulation section explicitly uses it, so the practical demonstration stands on its own.

The problem is Theorem 1. The proof reuses the RMPC's nominal trajectory as the performance horizon trajectory at the next step, but the performance horizon evolves under (2c), which includes the learned residual d. Assumption 2 only says d lies in W and Xhat_j contains Xbar_j. That does not imply Xbar_j + W is inside Xhat_j, so a state that is feasible for the performance horizon can be pushed out by the learned term. A simple scalar example (f=0.5x, u=0, W={3}, Xhat=Xbar=[-4,4], d=3) shows the claimed recursive feasibility fails: the performance state after one step is 4.5, violating the constraint. This is not a minor technicality—it is the advertised hard-constrained safety guarantee. The paper's own use of the soft-constrained version in the simulations means the practical claims are not directly affected, but the abstract and the title overstate what is proven.

The paper leans heavily on self-cited [16] and [17] for the implementation, but that is not a problem by itself; the issue is that the containment condition for Xhat is never made strong enough.

Who should read it: people working on safe learning and MPC will find the framework and the soft-constrained result useful, and the simulation is a decent illustration. But as submitted, the central theorem does not hold. I would send it to peer review—the framework deserves scrutiny and the soft-constrained result is worth publishing—but I would require a major revision: either strengthen Assumption 2 with a robust invariance condition on Xhat, or remove the hard-constrained claim and make Theorem 2 the centerpiece.

Referee Report

2 major / 3 minor

Summary. The paper proposes a contingency MPC (CMPC) framework that couples a robust MPC (RMPC) contingency horizon with a learning-based MPC (LB-MPC) performance horizon through a shared initial state and first control input. The authors claim that, under two assumptions on the RMPC and the learned disturbance model, the CMPC inherits the robust recursive feasibility of the RMPC (Theorem 1), and that a soft-constrained variant remains recursively feasible under the RMPC assumption alone (Theorem 2). The framework is instantiated with an RMPC from [16] and a Gaussian-process MPC from [17] and evaluated in a lane-merging simulation with a sensitivity study over 231 scenarios.

Significance. If valid, the framework would be a useful conceptual unification of contingency planning and safe learning, and the soft-constrained Theorem 2 does provide a correct safety guarantee for the specific implementation used in the simulations. The numerical study is more thorough than typical for a Letters paper and demonstrates the expected qualitative benefits: CMPC is more assertive than RMPC and safer than standalone GP-MPC. The main advertised hard-constrained inheritance result, however, is false as stated, so the paper's central theoretical claim needs substantial revision before the contribution can be accepted.

major comments (2)

[Section IV, Theorem 1 and its proof] The step 'we can select \hat x_{k+1|k+1} := \bar x_{k+1|k+1} and \hat U_{k+1} := \bar U_{k+1} that leads to a state sequence ... \in \hat X_j|k+1' is invalid. The sequence generated by (2c) includes the learned term d(\cdot,\cdot,D_{k+1}) and is not the RMPC nominal sequence generated by (2b). Assumption 2 gives only d\in W and \hat X_j|k \supseteq \bar X_j; it does not imply the robust one-step containment \hat X_{j+1}|k+1 \supseteq f(\hat X_j|k+1, \bar u_{j|k+1}) + W, which is what the argument would need. A concrete counterexample is the scalar system f(x,u)=0.5x, U={0}, W=[0,3], d=3, \hat X_j = \bar X_j = [-4,4], N=1, X={0}. At x_0=0 the OCP is feasible (\hat x_{1|0}=3 \in [-4,4]), but after w_0=3 the state is x_1=3 and the only candidate performance successor is f(3,0)+3=4.5 \notin [-4,4], so no feasible performance horizon exists at time 1. Thus Theorem 1's hard-constrained inheritance claim, which is also the unqualified claim in the abstract, is false under Assumptions 1 and 2.
[Section IV, Theorem 2 proof] The sentence 'It follows directly from the proof of Thm. 1 that \hat x_{k+1|k+1} := \bar x_{k+1|k+1} and \hat U_{k+1} := \bar U_{k+1} satisfy (5)' repeats the same non-sequitur, because the performance trajectory under (2c) is not the nominal trajectory. The theorem's conclusion is nonetheless recoverable: after fixing the candidate \hat x and \hat U, compute the predicted performance states from (2c) and choose each slack \epsilon_{k+1+j|k+1} large enough so that the corresponding predicted state lies in \hat X_j|k+1(\epsilon), which is possible because \hat X_j|k+1(\epsilon) \to R^{nx} as \epsilon \to \infty. The proof should be rewritten accordingly; since the simulation study and the closed-loop safety claims of Section VII rely on Theorem 2, this is not a purely stylistic point.

minor comments (3)

[Section VII-A] The maximum computation time of 2.517 s exceeds the sampling period T_s=0.25 s by an order of magnitude; the statement that real-time tractability is reachable with code optimization should be supported or qualified.
[Section VII-B, Table I] The text says RMPC and CMPC 'never exceed the safety constraint,' while CMPC has nonzero slack values (1e-7 and 1e-6); please clarify that these slacks are for the soft performance-horizon constraints and do not correspond to closed-loop safety violations.
[Section VII-B] The 12 and 11 excluded initial conditions for RMPC and CMPC should be mentioned in the conclusions as a limitation; the safety guarantee is conditional on reachability of the terminal invariant set.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reduction: Theorem 1 is a conditional implication, and the identified flaw is a proof gap, not a circularity.

full rationale

The central derivation is Theorem 1, which states that feasibility of the CMPC (2) at time k implies feasibility at k+1 under Assumptions 1 and 2. This is a conditional implication and not a restatement of either assumption: Assumption 1 concerns only the RMPC subproblem (3), and Assumption 2 bounds the learned residual and compares the tightened constraint sets, but neither assumption contains the CMPC feasibility conclusion. The paper's own equations (2b)-(2i) define both horizons, and the proof constructs a candidate from the RMPC solution; no parameter is fitted and then reported as a prediction. The GP-MPC implementation is used for prediction, not for the safety proof, and the citations [16] and [17] are self-citations that supply an implementation and approximation machinery, not the theorem itself. The reader-identified flaw, namely that the proof does not account for the d term in (2c) when reusing the RMPC trajectory so that Assumption 2 is too weak to keep the performance state inside Xhat, is a correctness gap in the proof rather than a circular reduction. The numerical section also openly reports infeasible initial states excluded from the KPIs, which is a stated limitation rather than a disguised fit. Overall, no equation in this paper is equivalent by construction to its own output.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central framework proof itself has no fitted constants, but the simulation and implementation rely on many hand-chosen parameters and imported components. The key assumptions are the robust recursive feasibility of the chosen RMPC and boundedness/containment of the learned model; the proof gap shows Assumption 2 is not sufficient. No new physical entities are introduced.

free parameters (6)

P = 0.5
In (16), weights robust cost against learning-based cost; chosen by authors, affects assertiveness.
rho = 1e4
Weight of l1 slack penalty in (5); chosen large to keep soft-constraint violations small.
GP kernel hyperparameters = sigma_d=0.7, L_d=diag(5,100,500,100)
GP kernel parameters in (13); set manually and control learned residual impact.
Number of GP inducing points M = M=4
Number of pseudo-inputs in sparse GP; approximation that affects prediction quality.
Prediction horizon N = N=20
Prediction horizon in (2); hand-chosen design parameter.
Cost weights Q, R, S = Q=10, R=1, S=10
Stage cost weights in (16); chosen by hand and affect trajectory aggressiveness.

assumptions (6)

domain assumption Underlying RMPC (3) is robustly recursively feasible (Assumption 1).
The proofs of Theorems 1 and 2 both invoke this to get a feasible robust plan at time k+1. It is imported from standard RMPC theory and, in the implementation, from the authors' own [16].
domain assumption Learned residual satisfies d(x,u,D_k) in W and performance constraint sets contain robust tightened sets: Xhat_j|k superset of Xbar_j (Assumption 2).
Used in Theorem 1 to reuse the robust solution for the performance horizon. It is not strong enough for the proof, which implicitly needs the performance trajectory under d to remain in Xhat.
domain assumption A minimizer of OCP (2) exists and is unique for every relevant state.
The framework assumes solvability to define the applied control law pi(x_k).
domain assumption The disturbance w_k is bounded in W with 0 in W and the system dynamics have the additive form (1).
Problem formulation in Sec. II; required for robust constraint tightening to be valid.
domain assumption The GP prediction model uses common approximations: joint Gaussianity, independent GP evaluations, first-order Taylor expansion, and sparse pseudo-inputs.
These approximations are needed for tractable multi-step GP-MPC and shape the Xhat_j|k sets in the implementation.
domain assumption The terminal sets Omega_1, Omega_2 and safety distance function D_safe from [16] provide robust control invariance.
The implementation's RMPC recursive feasibility rests on these imported sets; they are not constructed or verified in this paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Contingency Model Predictive Control Framework for Safe Learning." pith.science (2026). https://pith.science/paper/JP5OUPOA

@misc{pith2026250522776,
  author       = {Pith},
  title        = {Pith review of: A Contingency Model Predictive Control Framework for Safe Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JP5OUPOA}},
  note         = {Machine review of arXiv:2505.22776}
}

read the original abstract

This research introduces a multi-horizon contingency model predictive control (CMPC) framework in which classes of robust MPC (RMPC) algorithms are combined with classes of learning-based MPC (LB-MPC) algorithms to enable safe learning. We prove that the CMPC framework inherits the robust recursive feasibility properties of the underlying RMPC scheme, thereby ensuring safety of the CMPC in the sense of constraint satisfaction. The CMPC leverages the LB-MPC to safely learn the unmodeled dynamics to reduce conservatism and improve performance compared to standalone RMPC schemes, which are conservative in nature. In addition, we present an implementation of the CMPC framework that combines a particular RMPC and a Gaussian Process MPC scheme. A simulation study on automated lane merging demonstrates the advantages of our general CMPC framework.

Figures

Figures reproduced from arXiv: 2505.22776 by the authors.

**Figure 1.** A schematic of CMPC framework with circles indicating the admissible state constraint sets for each horizon. as is common in various RMPC approaches [16], [18]. The LB horizon uses the data set Dk to adapt the tightened state constraint set Xˆ j|k ⊆ X in (2i) over time k ∈ N. Finally, let (¯x ∗ k|k ,U¯ ∗ k ,Uˆ ∗ k ) be a minimizer of (2) for xk, which we assume exists, at time k ∈ N for cost function J. For simplici… view at source ↗

**Figure 2.** (i) Initially, the ego vehicle (Agent 1, gray vehicle) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 2.** Set-up of the lane-merging scenario. The objective is to use CMPC to control Agent 1, completing steps (i) to (iii) of the scenario, learning the acceleration policy g of Agent 2 and adhering to all the safety constraints. VI. CMPC FOR AUTOMATED LANE MERGING In this section, we apply the CMPC framework proposed in Sec. III to the automated lane merging problem (see Sec. V) by combining the RMPC from [16] for safety … view at source ↗

Figures from the paper (1 more)

**Figure 3.** Figure 3: Comparison of RMPC (left), CMPC (middle) and GP-MPC (right) for one lane merging scenario, see Sec. VII-A. Shown are the trajectories of the path coordinate s i , velocity v i , acceleration ui and relative distance |∆s| of Agent 1 and 2. The lower limit of |∆s| is imp…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

[11]

Contingency model predictive control for automated vehicles,

J.P. Alsterda, M. Brown, and J.C. Gerdes, “Contingency model predictive control for automated vehicles,” American Control Conference, 2019

work page 2019
[13]

Driver- in-the-loop contingency MPC with invariant sets,

K.S. Schweidel, S.M. Koehler, V .R. Desaraju, and M. Baric, “Driver- in-the-loop contingency MPC with invariant sets,” European Control Conference, 2022

work page 2022
[15]

Invariant safe contingency model predictive control for intersection coordination of mixed traffic,

X. Chen and J. M ˚artensson, “Invariant safe contingency model predictive control for intersection coordination of mixed traffic,” IEEE Intelligent Transportation Systems Conference, 2023

work page 2023
[16]

Automated Lane Merging with Safety Guarantees: A Robust Model Predictive Control Approach with Experimental Validation,

M.E. Geurts, A. Katriniok, E. Silvas, N.J. Brouwer and W.P.M.H. Heemels. “Automated Lane Merging with Safety Guarantees: A Robust Model Predictive Control Approach with Experimental Validation,” Techrxiv, 2025

work page 2025
[17]

Online Learning of Interaction Dynamics with Dual Model Predictive Control for Multi-Agent Systems Using Gaussian Processes

T.M.J.T. Baltussen, E. Lefeber, R. T ´oth, W.P.M.H. Heemels and A. Katriniok, “Online learning of interaction dynamics with dual model predictive control for multi-agent systems using Gaussian processes,” American Control Conference 2025, arXiv:2409.00432

work page Pith review arXiv 2025
[1]

A safe control architecture based on robust model predictive control for autonomous driving,

M. Nezami, N.T. Nguyen, G. Mannel, H.S. Abbas, and G. Schildbach, “A safe control architecture based on robust model predictive control for autonomous driving,” American Control Conference, 2022

work page 2022
[2]

Tractable robust MPC design based on nominal predictions,

I. Alvarado, P. Krupa, D. Limon, and T. Alamo, “Tractable robust MPC design based on nominal predictions,” J. Process Control, 2022

work page 2022
[3]

Hewing, J

L. Hewing, J. Kabzan and M. N. Zeilinger, ”Cautious Model Predictive Control Using Gaussian Process Regression,” in IEEE Transactions on Control Systems Technology, 2020.,

work page 2020

Show all 21 references

[4]

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 Systems, 2020

2020
[5]

Safe reinforcement learning using robust MPC,

M. Zanon and S. Gros, “Safe reinforcement learning using robust MPC,” IEEE Transactions on Automatic Control, 2021

2021
[6]

Wabersich and M.N

K.P. Wabersich and M.N. Zeilinger, ”Linear Model Predictive Safety Certification for Learning-Based Control,” IEEE Conference on Decision and Control, 2018

2018
[7]

Provably safe and robust learning-based model predictive control,

A. Aswani, H. Gonzalez, S.S. Sastry, and C. Tomlin, “Provably safe and robust learning-based model predictive control,” Automatica, 2013

2013
[8]

Stochastic model predictive control with a safety guarantee for automated driving,

T. Brudigam, M. Olbrich, D. Wollherr, and M. Leibold, “Stochastic model predictive control with a safety guarantee for automated driving,” IEEE Transactions on Intelligent Vehicles, 2023

2023
[9]

Robust model predictive shielding for safe reinforcement learning with stochastic dynamics,

S. Li and O. Bastani, “Robust model predictive shielding for safe reinforcement learning with stochastic dynamics,” IEEE ICRA, 2020

2020
[10]

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 and T.A. Badgwell, “Fusion of Machine Learning and MPC under Uncertainty: What Advances Are on the Horizon?,” American Control Conference, 2022

2022
[12]

Augmenting MPC schemes with active learning: Intuitive tuning and guaranteed performance,

R. Soloperto, J. Kohler, and F. Allgower, “Augmenting MPC schemes with active learning: Intuitive tuning and guaranteed performance,” IEEE Control Systems Letters, 2020

2020
[14]

Contingent nonlinear model predictive control for collision imminent steering in uncertain environments,

J. Dallas, J. Wurts, J. L. Stein, and T. Ersal, “Contingent nonlinear model predictive control for collision imminent steering in uncertain environments,” IFAC-PapersOnLine, 2020

2020
[18]

Model Predictive Control: Theory, Computation, and Design

J.B. Rawlings, D.Q. Mayne and M. Diehl, “Model Predictive Control: Theory, Computation, and Design”, Nob Hill Publishing, 2017

2017
[19]

Gaussian processes for machine learning

C.E. Rasmussen and C.K. Williams, “Gaussian processes for machine learning”, MIT Press, 2006

2006
[20]

On the implementation of an interior- point filter line-search algorithm for large-scale nonlinear programming,

A. W ¨achter and L. T. Biegler, “On the implementation of an interior- point filter line-search algorithm for large-scale nonlinear programming,” Mathematical Programming, 2006

2006
[21]

CasADi: a software framework for nonlinear optimization and optimal control,

J. A. E. Andersson, J. Gillis, G. Horn, J. B. Rawlings, and M. Diehl, “CasADi: a software framework for nonlinear optimization and optimal control,” Mathematical Programming Computation, 2019

2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [11]

Contingency model predictive control for automated vehicles,

J.P. Alsterda, M. Brown, and J.C. Gerdes, “Contingency model predictive control for automated vehicles,” American Control Conference, 2019

work page 2019

[2] [13]

Driver- in-the-loop contingency MPC with invariant sets,

K.S. Schweidel, S.M. Koehler, V .R. Desaraju, and M. Baric, “Driver- in-the-loop contingency MPC with invariant sets,” European Control Conference, 2022

work page 2022

[3] [15]

Invariant safe contingency model predictive control for intersection coordination of mixed traffic,

X. Chen and J. M ˚artensson, “Invariant safe contingency model predictive control for intersection coordination of mixed traffic,” IEEE Intelligent Transportation Systems Conference, 2023

work page 2023

[4] [16]

Automated Lane Merging with Safety Guarantees: A Robust Model Predictive Control Approach with Experimental Validation,

M.E. Geurts, A. Katriniok, E. Silvas, N.J. Brouwer and W.P.M.H. Heemels. “Automated Lane Merging with Safety Guarantees: A Robust Model Predictive Control Approach with Experimental Validation,” Techrxiv, 2025

work page 2025

[5] [17]

Online Learning of Interaction Dynamics with Dual Model Predictive Control for Multi-Agent Systems Using Gaussian Processes

T.M.J.T. Baltussen, E. Lefeber, R. T ´oth, W.P.M.H. Heemels and A. Katriniok, “Online learning of interaction dynamics with dual model predictive control for multi-agent systems using Gaussian processes,” American Control Conference 2025, arXiv:2409.00432

work page Pith review arXiv 2025

[6] [1]

A safe control architecture based on robust model predictive control for autonomous driving,

M. Nezami, N.T. Nguyen, G. Mannel, H.S. Abbas, and G. Schildbach, “A safe control architecture based on robust model predictive control for autonomous driving,” American Control Conference, 2022

work page 2022

[7] [2]

Tractable robust MPC design based on nominal predictions,

I. Alvarado, P. Krupa, D. Limon, and T. Alamo, “Tractable robust MPC design based on nominal predictions,” J. Process Control, 2022

work page 2022

[8] [3]

Hewing, J

L. Hewing, J. Kabzan and M. N. Zeilinger, ”Cautious Model Predictive Control Using Gaussian Process Regression,” in IEEE Transactions on Control Systems Technology, 2020.,

work page 2020