Pith. sign in

REVIEW 4 major objections 5 minor 35 references

Safety Meets Speed: Accelerated Neural MPC with Safety Guarantees and No Retraining

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that replacing a full CBF-MPC horizon with a learned, parameter-adaptive value function preserves safety while cutting online solve time by roughly 200x on embedded hardware.

desk verdict Useful engineering idea with a real speedup; the stability theorem is not bulletproof—Assumption 5 assumes the conclusion and Eq. (38b is wrong—but the paper deserves a serious referee for the embedded MPC angle. read the letter →

arxiv 2509.06404 v1 pith:7NFA2UCD submitted 2025-09-08 cs.RO

Kaikai Wang , Tianxun Li , Liang Xu , Qinglei Hu , Keyou You This is my paper · ORCID
classification cs.RO
keywords controlbarrierfunctionsmodelpredictiveneuralvaluefunctionparametersensitivityembeddedobstacleavoidanceDAGGERadaptive
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 sets out to break the speed-safety trade-off in embedded model predictive control. It proposes BAN-MPC, which replaces the long prediction horizon of a CBF-constrained MPC with a short horizon plus a neural network that approximates the remaining cost-to-go. A second neural network learns how that cost-to-go changes when model parameters drift, so the controller adapts without retraining. The authors claim this preserves the safety guarantees of the original MPC while cutting online computation by two orders of magnitude on embedded hardware.

What carries the argument

The load-bearing object is the parameter-adaptive neural value function V_BAN-MPC(x, theta) = V_NN(x) + grad V_NN(x)(theta - theta_nom), a first-order Taylor correction of a learned cost-to-go. It is embedded as a terminal cost in a short-horizon MPC (horizon M much smaller than N), while the CBF constraint (Delta H + gamma H >= 0) is kept intact in the online problem. The VF-DAGGER training loop (Algorithm 1) iteratively aggregates on-policy data so the learned value function tracks the true CBF-MPC value function despite distribution shift. The neural sensitivity network, trained on the Jacobian of the MPC value function at nominal parameters, supplies the adaptation term and avoids retrai

What would settle it

Measure, on the actual closed-loop runs, the ratio ||V_BAN(x_{k+1}, theta) - V_MPC(x_{k+1}, theta) - (V_BAN(x_k, theta) - V_MPC(x_k, theta))|| / |V_BAN(x_k, theta) - V_MPC(x_k, theta)|. If this ratio exceeds the chosen gamma at any state along a successful trajectory, Assumption 5 fails and the proof's premise collapses; alternatively, perturb parameters beyond 15% and observe whether the value-function error term diverges as the active-set-change caveat predicts.

Watch

Extended reading notes

Core claim

The central claim is that the optimal value function of a CBF-MPC problem can be learned offline and reused online as a terminal cost for a much shorter horizon, with a learned sensitivity term correcting it under parameter variation. With the CBF constraint kept in the short-horizon problem, strict obstacle-avoidance safety is retained by forward invariance; the neural value function only accelerates the solve. The paper proves probabilistic practical exponential stability of the closed loop (Theorem 4) and demonstrates on a Jetson Nano that the resulting BAN-MPC solves roughly 200 times faster than CBF-MPC, keeps trajectories within 5% control error for model parameter changes up to 15%, a

Load-bearing premise

The proof that BAN-MPC is practically exponentially stable assumes an unverified bound on how fast the value-function approximation error can grow along closed-loop trajectories (Assumption 5); if that bound does not hold, the convergence guarantee collapses even though the CBF constraint still preserves safety.

Editorial extensions

If this is right

  • Embedded robots—drones, rovers, and mobile platforms—could run safety-guaranteed MPC at real-time rates on inexpensive processors rather than needing full-size compute.
  • A single offline training pass covers a whole range of mass, friction, and geometric parameters (up to 15% deviation), eliminating retraining when hardware or payload changes.
  • The VF-DAGGER guarantee replaces the quadratic horizon-dependence of behavioral cloning with a sublinear dependence, making learned MPC value functions viable for long-horizon navigation.
  • The stability theorem (Theorem 4) gives explicit formulas for convergence rate and attraction radius in terms of network error and parameter deviation, so designers can budget how much approximation error the closed loop tolerates.

Reading between the lines

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

  • The 200x speedup bundles the effect of a shorter horizon with the neural value function. Isolating those two contributions would require a baseline short-horizon MPC without the learned cost, and the paper does not report solve times for that controller on the embedded board.
  • Assumption 5, the error-growth bound, is never checked empirically. A reader who cares about the formal guarantee should measure that ratio directly on HIL trajectories; if it crosses gamma, the advertised stability margin may be smaller than claimed even if CBF safety still holds.
  • The sensitivity correction is first-order, so its validity region is tied to the active-set stability of the underlying nonlinear program. Combining BAN-MPC with an online estimator that keeps theta within the nominal neighborhood would extend the 15% envelope to drifting or unobservable parameters, a direction the authors list as future work.
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 BAN-MPC, a short-horizon MPC whose terminal cost is an offline-learned neural value function augmented by a neural sensitivity term, with safety enforced by Control Barrier Function constraints. The authors report hardware-in-the-loop experiments on a Jetson Nano for unicycle and quadrotor obstacle-avoidance tasks, showing large computational speedups over CBF-MPC, high domain/boundary safety, and graceful performance under parameter perturbations up to 15%. The theoretical appendix claims a probabilistic practical exponential stability guarantee for the BAN-MPC closed loop. The central empirical claims are plausible and the experiments are well structured, but the theoretical proof, as written, contains an algebraic error and relies on an unproved, effectively circular assumption about the evolution of the value-function approximation error. The advertised stability guarantee is therefore not established.

Significance. If the stability theorem were valid, the paper would make a meaningful contribution: it targets an important practical bottleneck (embedded MPC) and combines neural value-function approximation, sensitivity-based adaptation, and CBF safety in a single framework. The HIL experiments are a real strength and support the claims about speed, safety, and adaptability in the specific scenarios tested. The VF-DAGGER procedure is a reasonable contribution and the finite-sample analysis follows known imitation-learning arguments. However, the theoretical part overreaches: the proof of Theorem 4 is conditional on an unverified growth condition on the value-function error, and a core algebraic step misstates the error decomposition. The framework may still be valuable as an empirical method with a conditional stability result, but the paper's current headline guarantee needs substantial revision.

major comments (4)
  1. [Appendix B, Eq. (38b)] Assumption 5 is the engine of Theorem 4: the convergence rate lambda and the attraction radius r are direct functions of gamma, and all later Lyapunov inequalities rely on it. Yet no derivation, reference, or experiment justifies the bound ||phi(x_{k+1},theta)-phi(x_k,theta)|| <= gamma ||V_BAN-MPC(x_k,theta)-V_MPC(x_k,theta)||. Theorems 2-3 give statistical guarantees on the surrogate policy performance J(pi)-J(pi*) and on one-step average losses; they do not imply a pointwise, along-trajectory contraction condition on the error function phi. In fact, the assumption postulates that the approximation error decays no slower than its current magnitude—essentially the benign error behavior practical stability needs to establish. The theorem should be rephrased as conditional on this assumption, or Assumption 5 should be proved from the learning guarantees with explicit constants.
  2. [Appendix B, Eq. (38b)] Equation (38b) is algebraically incorrect. From V_BAN-MPC = V_MPC + phi, the correct identity is Delta V_BAN-MPC = Delta V_MPC + (phi(x_{k+1}) - phi(x_k)). The manuscript instead writes Delta V_BAN-MPC = Delta V_MPC + ||phi(x_{k+1}) - phi(x_k)||. Replacing the signed difference by its norm is invalid unless the difference is known to be nonnegative, and no such monotonicity is established. This invalidates the subsequent inequality chain (38c)-(38d), which is load-bearing for the Lyapunov decrease argument. The proof needs a one-sided bound on the signed term or a different decomposition.
  3. [Assumption 3 vs. Eq. (39)] Assumption 3 states that the nominal MPC value function satisfies Delta V_MPC(x, theta_nom) <= -c||x||^2, which is normally a property of the value function along the optimal nominal MPC closed-loop policy. However, in the proof of Theorem 4, Eq. (39) applies this bound to the state trajectory x_k generated by the BAN-MPC controller. The BAN-MPC control input is not the expert MPC input; it is the solution to (13) with a much shorter horizon and a neural terminal cost. No argument shows that the nominal V_MPC decreases at rate c along BAN-MPC trajectories. This is a second hidden assumption that is load-bearing for the Lyapunov proof.
  4. [Eq. (17)/(13a) and Theorem 4] The quantity V_BAN-MPC defined in Eq. (17) is the parameter-adaptive neural terminal cost, not the optimal value function of the BAN-MPC optimization problem (13) and not the cost-to-go of the closed-loop system. Nevertheless, Theorem 4 uses Delta V_BAN-MPC along closed-loop states as a Lyapunov function (Eq. (38a)-(40c)). For the argument to be valid, V_BAN-MPC must be a Lyapunov candidate for the closed loop, which requires it to dominate the true closed-loop cost-to-go and to decrease under the BAN-MPC policy. These properties are not proved and do not follow from definition (17). The theorem as stated conflates the terminal-cost surrogate with the closed-loop value function; the authors should either prove the necessary Lyapunov properties or restrict the claim to the surrogate function.
minor comments (5)
  1. [Definition 1] The condition in Definition 1 writes 'H(x_0) >= 0 and there exists a control input u_k...' but H(x_0) is never quantified over states. It should be stated for all x_0 in S, or the intended initial-state condition should be made explicit.
  2. [Appendix B, Theorem 2] The proof of Theorem 2 assumes a constant u bounding the single-step action cost difference, but this constant is not included in the theorem statement or in the overview of assumptions. Please state all constants and their dependencies in the theorem.
  3. [Appendix B, Eq. (43)] The algebraic step leading to Eq. (43) is difficult to follow because the constant D is introduced in Step 5 without a precise relation to the earlier Gamma_V and Gamma_theta terms. Please define D explicitly and re-derive the state-norm bound step by step.
  4. [Appendix B, Step 6] The phrase 'As T -> infinity' is confusing because T in the learning theorems is the task horizon, while the stability theorem is about discrete-time indices k. The relation between the task horizon T and the closed-loop time k should be clarified, or the limiting statement removed.
  5. [General] The manuscript does not mention any plan to release code or datasets. For reproducibility of the HIL numbers and for verification of the claimed constants, including the training data generation procedure and the neural network training details would be helpful.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: speed/safety claims benchmarked externally; stability theorem is conditional on stated assumptions, not self-referential.

full rationale

The paper's central empirical claims are comparisons against external baselines (CBF-MPC, short-horizon MPC, AMPC) on a Jetson Nano, so the reported 200x speedup, suboptimality, and safety metrics are not fitted inputs renamed as predictions. The theoretical stability result (Theorem 4) depends on Assumption 5, which bounds the trajectory variation of the value-function approximation error by gamma times its current magnitude: ||phi(x_{k+1},theta)-phi(x_k,theta)|| <= gamma ||V_BAN-MPC(x_k,theta)-V_MPC(x_k,theta)||. This is an explicit hypothesis, not a conclusion proven from earlier results; the paper gives no verification or derivation of it. That is a rigor gap / omitted proof, not circularity: if the assumption held, the Lyapunov argument would go through, and the assumption is not equivalent to the desired stability statement. Eq. (38b) also contains an algebraic error (it writes an equality with a norm where a signed difference belongs), but the subsequent inequality can be read as an upper bound, so it is a proof defect rather than a self-referential definition. The only self-citation, [4], appears in Related Work and is not load-bearing. VF-DAGGER's guarantees are imported from the external DAGGER analysis of Ross and Bagnell. The CBF constraint (13f) enforces safety online by construction, so the 100% safety numbers are a consequence of the constraint, not a circular prediction. Overall, no derivation step reduces to its own input; the main concerns are correctness/rigor, not circularity.

Assumptions & free parameters 6 free parameters · 7 assumptions · 0 invented entities

The paper introduces no new physical entities; the additional objects are fitted neural networks, which are captured as free parameters. The main unstated load is Assumption 5, which is ad hoc to the paper.

free parameters (6)
  • CBF decay rate gamma = 0.3
    Chosen by hand in all experiments (Section IV-A); determines the CBF safety condition (7f) and affects the feasible set.
  • V_NN network weights = trained, no checkpoint released
    Learned by MSE regression on CBF-MPC value data (Eq. 15); the approximation quality is central to BAN-MPC's performance.
  • d(V_NN) network weights = trained, no checkpoint released
    Learned to match the Jacobian d(V_MPC)/d(theta) at theta_nom; this fitted sensitivity is the mechanism for retraining-free adaptation (Eq. 17).
  • Softmin parameter rho = not reported
    Smooths the obstacle-safety function in Eq. (14); chosen by hand but no value is given, so experiments are not fully reproducible.
  • VF-DAGGER mixing weight beta_i = not reported
    Tunes the iterative blend in Eq. (16); no schedule or value is specified in the experiments.
  • Stability constants alpha1, alpha2, c, L, L_V, eps_V, eps_grad, M_theta = not estimated
    Appear in Theorem 4's rate and radius but are never computed for the unicycle or quadrotor, so the theoretical guarantee is not made quantitative.
assumptions (7)
  • domain assumption Cost and constraints are twice continuously differentiable near the KKT point, and LICQ, SOSC, and strict complementarity hold (Section II-B).
    Required for the implicit-function sensitivity result (12) used to justify the value-function sensitivity network.
  • domain assumption For all parameter perturbations in a neighborhood, the active constraint set is unchanged (Section II-B).
    Needed for the first-order prediction s*(theta+delta theta) to be valid; the experiments only vary parameters within 15% but active-set invariance is not checked.
  • ad hoc to paper The nominal MPC value function satisfies alpha1||x||^2 <= V_MPC(x,theta_nom) <= alpha2||x||^2 and Delta V_MPC <= -c||x||^2 (Appendix B).
    Assumed without proof for a finite-horizon receding-horizon MPC with CBF constraints; this strong Lyapunov condition is not generally implied by MPC stability.
  • domain assumption The neural network approximation errors e_V and e_grad are bounded with probability 1-delta (Appendix B).
    Standard approximation-theoretic assumption, but the bounds eps_V and eps_grad are not estimated for the trained networks.
  • ad hoc to paper ||phi(x_{k+1},theta)-phi(x_k,theta)|| <= gamma ||V_BAN-MPC(x_k,theta)-V_MPC(x_k,theta)|| (Appendix B, Assumption 5).
    Unproved growth condition on the value-function error along trajectories; the stability theorem and its constants are derived directly from it, making it the weakest load-bearing assumption.
  • domain assumption The real parameters stay within ||theta-theta_nom|| <= M_theta and the active set remains unchanged (Appendix B, Assumption 6).
    Bounded parameter deviation is needed for the first-order adaptation Eq. (17); the experiments use <=15% but no guard is given for when this fails.
  • domain assumption The parameter deviation theta-theta_nom is available online.
    Used implicitly in Eq. (17); in real systems parameters such as mass or friction must be estimated, which the paper leaves to future work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Safety Meets Speed: Accelerated Neural MPC with Safety Guarantees and No Retraining." pith.science (2026). https://pith.science/paper/7NFA2UCD

@misc{pith2026250906404,
  author       = {Pith},
  title        = {Pith review of: Safety Meets Speed: Accelerated Neural MPC with Safety Guarantees and No Retraining},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7NFA2UCD}},
  note         = {Machine review of arXiv:2509.06404}
}
read the original abstract

While Model Predictive Control (MPC) enforces safety via constraints, its real-time execution can exceed embedded compute budgets. We propose a Barrier-integrated Adaptive Neural Model Predictive Control (BAN-MPC) framework that synergizes neural networks' fast computation with MPC's constraint-handling capability. To ensure strict safety, we replace traditional Euclidean distance with Control Barrier Functions (CBFs) for collision avoidance. We integrate an offline-learned neural value function into the optimization objective of a Short-horizon MPC, substantially reducing online computational complexity. Additionally, we use a second neural network to learn the sensitivity of the value function to system parameters, and adaptively adjust the neural value function based on this neural sensitivity when model parameters change, eliminating the need for retraining and reducing offline computation costs. The hardware in-the-loop (HIL) experiments on Jetson Nano show that BAN-MPC solves 200 times faster than traditional MPC, enabling collision-free navigation with control error below 5\% under model parameter variations within 15\%, making it an effective embedded MPC alternative.

Figures

Figures reproduced from arXiv: 2509.06404 by the authors.

Figure 1
Figure 1. Overview of BAN-MPC. The framework is divided into an offline phase and an online phase. In the offline phase, training data is generated using CBF-MPC, and the neural value function and neural sensitivity are learned using neural networks. In the online phase, the neural value function and neural sensitivity are integrated into the optimization objective of Short-horizon MPC, with CBF ensuring strict safety guarant… view at source ↗
Figure 3
Figure 3. Hardware-in-the-Loop Evaluation System Diagram. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Obstacle avoidance navigation task for the unicycle. The gray [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Obstacle avoidance navigation task for the unicycle with a changed [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Quadcopter obstacle avoidance navigation trajectory and projection plot. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Quadcopter obstacle avoidance navigation trajectory and projection plot after modifying mass parameters. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: BAN-MPC’s parameter adaptive performance validation [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Dynamic obstacle-avoidance task using BAN-MPC [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 34 canonical work pages

  1. [13]

    On building myopic mpc policies using supervised learning,

    C. A. Orrico, B. Yang, and D. Krishnamoorthy, “On building myopic mpc policies using supervised learning,” 2024. [Online]. Available: https://arxiv.org/abs/2401.12546

    work page arXiv 2024

  2. [22]

    Parameter-Adaptive Approximate MPC: Tuning Neural-Network Controllers without Retraining

    H. Hose, A. Gr ¨afe, and S. Trimpe, “Parameter-adaptive approximate mpc: Tuning neural-network controllers without retraining,” 2024. [Online]. Available: https://arxiv.org/abs/2404.05835

    work page Pith review arXiv 2024

  3. [1]

    A convergent dynamic window approach to obstacle avoidance,

    P. Ogren and N. Leonard, “A convergent dynamic window approach to obstacle avoidance,”IEEE Trans. Rob., vol. 21, no. 2, pp. 188–195, 2005

    work page 2005

  4. [2]

    Augmenting reinforcement learning with behavior primitives for diverse manipulation tasks,

    S. Nasiriany, H. Liu, and Y . Zhu, “Augmenting reinforcement learning with behavior primitives for diverse manipulation tasks,”Proc. IEEE Int. Conf. Robot. Autom. (ICRA), pp. 7477–7484, 2022

    work page 2022

  5. [3]

    Online trajectory planning for multiple quadrotors in dynamic environments using relative safe flight corridor,

    J. Park and H. J. Kim, “Online trajectory planning for multiple quadrotors in dynamic environments using relative safe flight corridor,” IEEE Robot. Autom. Lett., vol. 6, no. 2, pp. 659–666, 2021

    work page 2021

  6. [4]

    Safe reinforcement learning and adaptive optimal control with applications to obstacle avoidance problem,

    K. Wang, C. Mu, Z. Ni, and D. Liu, “Safe reinforcement learning and adaptive optimal control with applications to obstacle avoidance problem,”IEEE Trans. Autom. Sci. Eng., vol. 21, no. 3, pp. 4599–4612, 2024

    work page 2024

  7. [5]

    Model-based constrained reinforcement learning using generalized control barrier function,

    H. Ma, J. Chen, S. Eben, Z. Lin, Y . Guan, Y . Ren, and S. Zheng, “Model-based constrained reinforcement learning using generalized control barrier function,”Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., pp. 4552–4559, 2021

    work page 2021

  8. [6]

    A real-time gp based mpc for quadcopters with unknown disturbances,

    N. Schmid, J. Gruner, H. S. Abbas, and P. Rostalski, “A real-time gp based mpc for quadcopters with unknown disturbances,”Proc. Am. Control Conf. (ACC), pp. 2051–2056, 2022

    work page

Show all 35 references
  1. [7]

    Real-time neural mpc: Deep learning model predictive control for quadrotors and agile robotic platforms,

    T. Salzmann, E. Kaufmann, J. Arrizabalaga, M. Pavone, D. Scara- muzza, and M. Ryll, “Real-time neural mpc: Deep learning model predictive control for quadrotors and agile robotic platforms,”IEEE Robot. Autom. Lett., vol. 8, no. 4, pp. 2397–2404, 2023

    2023

  2. [8]

    Clf-cbf constraints for real-time avoidance of multiple obstacles in bipedal locomotion and navigation,

    J. Liu, M. Li, J. W. Grizzle, and J.-K. Huang, “Clf-cbf constraints for real-time avoidance of multiple obstacles in bipedal locomotion and navigation,”Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), pp. 10 497–10 504, 2023

    2023

  3. [9]

    Control barrier function based quadratic programs for safety critical systems,

    A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic programs for safety critical systems,”IEEE Trans. Autom. Control, vol. 62, no. 8, pp. 3861–3876, 2017

    2017

  4. [10]

    Safety barrier certificates for path integral control: Safety-critical control of quadrotors,

    T. Jin, J. Di, X. Wang, and H. Ji, “Safety barrier certificates for path integral control: Safety-critical control of quadrotors,”IEEE Robot. Autom. Lett., vol. 8, no. 9, pp. 6006–6012, 2023

    2023

  5. [11]

    Multi-layered safety for legged robots via control barrier functions and model predictive control,

    R. Grandia, A. J. Taylor, A. D. Ames, and M. Hutter, “Multi-layered safety for legged robots via control barrier functions and model predictive control,”Proc. IEEE Int. Conf. Robot. Autom., pp. 8352– 8358, 2021

    2021

  6. [12]

    Near-optimal rapid mpc using neural networks: A primal-dual policy learning framework,

    X. Zhang, M. Bujarbaruah, and F. Borrelli, “Near-optimal rapid mpc using neural networks: A primal-dual policy learning framework,” IEEE Trans. Control Syst. Technol., vol. 29, no. 5, pp. 2102–2114, 2021

    2021

  7. [14]

    Constraints- informed neural-laguerre approximation of nonlinear mpc with application in power electronics,

    D. Xu, R. Aerts, P. Karamanakos, and M. Lazar, “Constraints- informed neural-laguerre approximation of nonlinear mpc with application in power electronics,” 2024. [Online]. Available: https: //arxiv.org/abs/2409.09436

    2024 arXiv

  8. [15]

    Approximating explicit model predictive control using constrained neural networks,

    S. Chen, K. Saulnier, N. Atanasov, D. D. Lee, V . Kumar, G. J. Pappas, and M. Morari, “Approximating explicit model predictive control using constrained neural networks,”Proc. Am. Control Conf., pp. 1520– 1527, 2018

    2018

  9. [16]

    Safe imitation learning of nonlinear model predictive control for flexible robots,

    S. Mamedov, R. Reiter, S. M. B. Azad, R. Viljoen, J. Boedecker, M. Diehl, and J. Swevers, “Safe imitation learning of nonlinear model predictive control for flexible robots,” 2024. [Online]. Available: https://arxiv.org/abs/2212.02941

    2024 arXiv

  10. [17]

    Bridging multi-task learning and meta- learning: Towards efficient training and effective adaptation,

    H. Wang, H. Zhao, and B. Li, “Bridging multi-task learning and meta- learning: Towards efficient training and effective adaptation,”Proc. Int. Conf. Mach. Learn., pp. 10 991–11 002, 2021

    2021

  11. [18]

    An evidential multi-target domain adaptation method based on weighted fusion for cross-domain pattern classification,

    L. Huang, W. Zhao, Y . Liu, D. Yang, A. W.-C. Liew, and Y . You, “An evidential multi-target domain adaptation method based on weighted fusion for cross-domain pattern classification,”IEEE Trans. Neural Netw. Learn. Syst., vol. 35, no. 10, pp. 14 218–14 232, 2024

    2024

  12. [19]

    Large scale model predictive control with neural networks and primal active sets,

    S. W. Chen, T. Wang, N. Atanasov, V . Kumar, and M. Morari, “Large scale model predictive control with neural networks and primal active sets,”Automatica, vol. 135, p. 109947, 2022

    2022

  13. [20]

    Model-free reinforce- ment learning for robust locomotion using demonstrations from tra- jectory optimization,

    M. Bogdanovic, M. Khadiv, and L. Righetti, “Model-free reinforce- ment learning for robust locomotion using demonstrations from tra- jectory optimization,”Frontiers in Robotics and AI, vol. 9, 2022

    2022

  14. [21]

    Sensitivity analysis for nonlinear programming using penalty methods,

    A. V . Fiacco, “Sensitivity analysis for nonlinear programming using penalty methods,”Math. Programm., vol. 10, pp. 287–311, 1976

    1976

  15. [23]

    Optimization-based collision avoidance,

    X. Zhang, A. Liniger, and F. Borrelli, “Optimization-based collision avoidance,”IEEE Trans. Control Syst. Technol., vol. 29, no. 3, pp. 972–983, 2021

    2021

  16. [24]

    A reduction of imitation learning and structured prediction to no-regret online learning,

    S. Ross, G. Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,”Proc. 14th Int. Conf. Artif. Intell. Statist. Workshop Conf. Proc., pp. 627–635, 2011

    2011

  17. [25]

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

    J. A. Andersson, J. Gillis, G. Horn, J. B. Rawlings, and M. Diehl, “Casadi: a software framework for nonlinear optimization and optimal control,”Math. Prog. Comp., vol. 11, pp. 1–36, 2019

    2019

  18. [26]

    Domain adaptive safety filters via deep operator learning,

    L. Manda, S. Chen, and M. Fazlyab, “Domain adaptive safety filters via deep operator learning,” 2024. [Online]. Available: https://arxiv.org/abs/2410.14528

    2024 arXiv

  19. [27]

    Data-driven mpc for quadrotors,

    G. Torrente, E. Kaufmann, P. F ¨ohn, and D. Scaramuzza, “Data-driven mpc for quadrotors,”IEEE Robot. Autom. Lett., vol. 6, no. 2, pp. 3769–3776, 2021

    2021

  20. [28]

    Efficient reductions for imitation learning,

    S. Ross and D. Bagnell, “Efficient reductions for imitation learning,” Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, vol. 9, pp. 661–668, 13–15 May 2010

    2010

  21. [29]

    Sensitivity analysis for nonlinear programming using penalty methods,

    A. V . Fiacco, “Sensitivity analysis for nonlinear programming using penalty methods,”Mathematical programming, vol. 10, no. 1, pp. 287– 311, 1976. APPENDIX A. Performance Guarantees of the VF-DAGGER Algorithm Traditional behavioral cloning trains policies by minimiz- ing a su...

    1976

  22. [30]

    Increasing the training epochs nD and trajectories per epochlyields the error decayε V = O 1 T andε ∇ =O 1 T

    Neural Network Approximation Error Convergence By Theorem 2 and Theorem 3 of VF-DAGGER, there exists a policyπsuch that the single-step expected errors satisfy∥e V (x)∥ ≤uT ε=O(1), and∥e ∇(x)∥ ≤u∇T ε= O(1)with probability1−δ. Increasing the training epochs nD and trajectories ...

  23. [31]

    Value Function Error Bound From [29] on MPC sensitivity, there exists a constant LV >0such that: ∥VMPC(x, θnom) + ∂ ∂θ VMPC(x, θ) θnom (θ−θ nom) −VMPC(x, θ)∥ ≤LV ||θ−θ nom||(36) Within the eventE={ |e V | ≤εV ∩ ∥e∇∥ ≤ε∇ }, the value function error between BAN-MPC and the origi...

  24. [32]

    Lyapunov Difference Decomposition Define error functionϕ(x, θ) =V BAN-MPC(x, θ)− VMPC(x, θ),∆V BAN-MPC(xk) =V BAN-MPC(xk+1)− VBAN-MPC(xk), and∆V MPC(xk) =V MPC(xk+1)−V MPC(xk). Then: ∆VBAN-MPC(xk, θ) =VBAN-MPC(xk+1, θ)−VBAN-MPC(xk, θ) (38a) = ∆VMPC(xk, θ) +∥ϕ(xk+1, θ)−ϕ(xk, θ)...

  25. [33]

    When∥x k∥> r: ∆VBAN-MPC ≤ −c∥xk∥2 + ΓV + Γθ (41a) <−cr 2 + ΓV + Γθ (41b) =−c ΓV + Γθ c−γα 2 + ΓV + Γθ (41c) = (ΓV + Γθ) −γα2 c−γα 2 (41d) <0(41e)

    Stability Condition Construction Define attraction radiusr= q ΓV +Γθ c−γα2 (γ∈ 0, c α2 ). When∥x k∥> r: ∆VBAN-MPC ≤ −c∥xk∥2 + ΓV + Γθ (41a) <−cr 2 + ΓV + Γθ (41b) =−c ΓV + Γθ c−γα 2 + ΓV + Γθ (41c) = (ΓV + Γθ) −γα2 c−γα 2 (41d) <0(41e)

  26. [34]

    The Lyapunov function evolution obeys:V BAN-MPC(xk+1, θ)≤ VBAN-MPC(xk, θ)−λ(Γ V + Γ θ), withλ= γα2 c−γα2 > 0, and further:∆V BAN-MPC ≤ −λα1∥xk∥2 +λD

    Exponential Decay Construction The value function satisfies two-sided inequalities: α1∥x∥2 −D≤V BAN-MPC(x, θ)≤α 2∥x∥2 +D, where D=ε V +(ε ∇ +L V )Mθ +LM θ is the total error bound. The Lyapunov function evolution obeys:V BAN-MPC(xk+1, θ)≤ VBAN-MPC(xk, θ)−λ(Γ V + Γ θ), withλ= γ...

  27. [35]

    State Norm Bound Combining the Lyapunov bounds, the state norm satisfies: α1∥xk∥2 −D≤β k(α2∥x0∥2 +D) + Dα2 α1 ⇒ ∥x k∥2 ≤ α2 α1 βk∥x0∥2 | {z } exponential term + D α1 βk + Dα2 α2 1 + D α1| {z } bias term (43) Letκ= q α2 α1 ,β k = 1− λα1 α2 k ≤e −λk. AsT→ ∞, the bias term isO(r ...

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.