Pith. sign in

REVIEW 5 major objections 6 minor 29 references

AI-Augmented Model Predictive Control for Safe and Adaptive Rendezvous and Proximity Operations

T0 review · 5 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read An adaptive layer that tunes the weights of a model-predictive controller — rather than learning thrust commands directly — raises rendezvous success in an adversarial satellite-capture setting, from 60 percent for a fixed controller to 72

desk verdict Useful engineering architecture with honest disclaimers, but the central adaptive-vs-fixed claim rests on a confounded comparison and thin statistics; worth refereeing with a request for equal tuning budgets and held-out seeds. read the letter →

arxiv 2607.16630 v1 pith:XZVPKX2R submitted 2026-07-18 cs.RO

Luca Sportelli , Tyler Barr , Cagri Kilic , Di Wu This is my paper
classification cs.RO
keywords adaptivemodelpredictivecontrolrendezvousandproximityoperationsClohessy-Wiltshiredynamicskeep-out-zoneconstraintsbarrierfunctionsupervisoryparametertuningadversarialmulti-agentspacecraftguidanceMonteCarlorobustnessevaluation
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

Rendezvous and proximity operations in adversarial orbit require balancing pursuit, safety, and adaptability. The paper proposes a two-loop controller: a constrained model-predictive controller handles low-level guidance, while an outer supervisory layer adjusts physically interpretable MPC parameters — tracking weights, keep-out-zone penalties, minimum-separation objectives — rather than replacing the controller with a learned policy. In the official Capture-the-Satellite competition environment, the adaptive MPC succeeded in 72 percent of comparison runs versus 60 percent for a fixed-parameter MPC with the same safety features and 28 percent for a naive MPC; a separate 1000-run campaign reported 94 percent success. The significance is that adaptive behavior can come from tuning an already constraint-aware optimization loop, preserving safety, interpretability, and real-time feasibility. The paper frames robustness as empirically validated, stating formal stability guarantees remain outside its scope.

What carries the argument

The load-bearing mechanism is a two-loop architecture. The inner loop is a constrained receding-horizon MPC: a convex quadratic program with Clohessy–Wiltshire linearized relative dynamics, actuator limits, predictive keep-out-zone constraints (with slack-variable softening), and an optional control-barrier-function projection. The outer loop is a supervisory layer that updates a bounded parameter vector P_k — position/velocity tracking weights, KOZ safety weight, minimum-separation objective, horizons, terminal weighting — through offline Cross-Entropy optimization of expected episode cost and online geometry-driven rules (proximity-based minimum-separation inflation, phase-dependent tracki

What would settle it

Re-run the 25-run comparison with a fixed-parameter controller whose parameters are optimized by the same derivative-free search (CEM) over the same bounded parameter set and the same seed distribution; if its success rate reaches or exceeds the adaptive controller's 72%, the paper's central claim of an online-adaptation benefit is contradicted. Also inspect slack activation and KOZ violation counts to confirm the adaptive advantage is not due to more frequent constraint relaxation.

Watch

Extended reading notes

Core claim

The paper claims that an outer-loop supervisory layer — which adjusts physically interpretable MPC parameters (tracking weights, keep-out-zone penalties, minimum-separation objectives) based on offline closed-loop optimization and online interaction geometry — improves closed-loop robustness and rendezvous performance in an adversarial satellite-capture scenario, while the inner constrained receding-horizon MPC, its linearized orbital dynamics model, actuator limits, and predictive safety constraints remain unchanged. In the authors' experiments, the adaptive controller achieved a 72% success rate over 25 comparison runs, versus 60% for a fixed-parameter MPC with the same safety handling and

Load-bearing premise

The headline comparison assumes the manually tuned fixed-parameter controller received tuning effort comparable to the CEM-optimized adaptive controller; if the fixed controller's 60% success rate reflects weaker tuning rather than the lack of adaptation, the central claim collapses — and the paper explicitly leaves formal stability guarantees out of scope.

Editorial extensions

If this is right

  • Adaptive robustness can be achieved without abandoning a constrained optimization controller, giving safety-critical orbital guidance a middle path between fixed MPC and end-to-end learning.
  • Because the supervised variables are physically meaningful, controller behavior can be audited after a mission: parameter trajectories correlate with orbital geometry and safety events, aiding certification-style review.
  • The architecture preserves real-time feasibility: the inner loop remains a small convex QP, and the outer loop requires only simple geometric updates.
  • Predictive keep-out-zone constraints, slack-variable feasibility handling, and the control-barrier-function filter remain active through the adaptive mechanism, so adaptation cannot bypass safety limits.
  • The same offline-tuned MPC core could be deployed with different online adaptation rules, separating controller validation from strategy tuning.

Reading between the lines

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

  • Editor inference: the 12-point success-rate advantage over the fixed controller may partly reflect tuning budget, since the adaptive controller's parameters were optimized offline with CEM while the fixed controller was manually tuned; the paper does not run a same-budget tuning control.
  • Editor inference: because the offline P* search and the evaluation draws from the same competition seed distribution, the 94% robustness figure is likely in-distribution; the paper does not claim out-of-distribution performance.
  • Editor inference: the bounded-parameter, soft-constraint architecture invites extensions such as online Bayesian or meta-learning of MPC weights; the paper does not propose those.
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

5 major / 6 minor

Summary. The paper proposes an adaptive Model Predictive Control (MPC) framework for spacecraft rendezvous and proximity operations in adversarial multi-agent environments. The architecture combines a constrained receding-horizon QP based on Clohessy–Wiltshire dynamics with predictive keep-out-zone constraints, slack-variable feasibility handling, and an optional Control Barrier Function filter. An outer supervisory layer modifies interpretable MPC parameters (tracking weights, horizons, safety penalties, minimum-separation objectives) using offline Cross-Entropy Method optimization of a base parameter vector and online geometry-driven rules. The framework is evaluated in the KSPDG Capture-the-Satellite competition environment, reporting a 94% success rate over 1,000 adaptive-MPC runs and a three-way comparison over 25 runs per configuration (Adaptive 72%, Fixed 60%, Naive 28%). The paper claims improved closed-loop robustness, adaptive maneuvering, and rendezvous performance relative to fixed-parameter MPC while preserving constraint-aware operation and interpretability.

Significance. If the central claim were conclusively supported, the contribution would be valuable: a modular, interpretable supervisory layer that tunes MPC parameters online while retaining the constrained optimization structure, rather than replacing it with an end-to-end learned policy. The MPC formulation is standard and internally consistent, and the paper is unusually explicit about the lack of formal recursive-feasibility and stability guarantees, appropriately framing the work as an engineering-oriented robustness study. The interpretability analysis linking adaptive parameters to physical safety geometry is also a strength. However, the empirical evidence for the headline claim is currently weak because the adaptive-vs-fixed comparison is confounded by unequal tuning effort and in-sample offline optimization, and because the reported sample sizes are too small to support the stated differences.

major comments (5)
  1. [§VI.B / Table 2 / §IV.E.1 (Eq. 24)] The Adaptive Agent's base parameter vector P* is obtained by minimizing expected episode return over the KSPDG seed distribution (Eq. 24), whereas the Fixed Agent is described only as 'manually tuned' with no tuning budget, search procedure, or parameter values. The reported 72% vs 60% success gap therefore conflates the online adaptive mechanism with unequal offline tuning effort or in-sample fitting to the evaluation distribution. To support the causal claim that the supervisory adaptation layer improves performance, the Fixed baseline should be given the same CEM tuning budget over the same parameter set (without online modulation), and both configurations should be evaluated on held-out seeds.
  2. [§V.C / §VI.B] The controller comparison uses 25 runs per configuration with no confidence intervals, error bars, or significance tests. For binary success with n=25, the standard error of the difference between 72% and 60% is approximately 13 percentage points, so the observed 12-point gap is not statistically significant. Furthermore, the Adaptive Agent achieves 94% success in the 1,000-run campaign but only 72% in the 25-run comparison, suggesting the comparison subset is not representative or the selection protocol is unstated. The paper should report confidence intervals, increase the number of runs, and explicitly describe how the 25 comparison runs were selected relative to the 1,000-run campaign.
  3. [§IV.E.2] The online adaptation constants — the 1.25 and 0.08 inflation factors, the phase-dependent tracking multipliers (4,25), (8,50), (12,80), (2.5,12), and the KOZ relaxation levels (0.5x, 0.3x, 0) — are presented as design choices. If any of these were selected or tuned using the KSPDG distribution, they inherit the same in-sample-fit concern as Eq. 24. The manuscript should state clearly which constants are fixed a priori, which were tuned, and whether the Fixed baseline had access to an equivalent tuning budget for comparable design choices.
  4. [§IV.D / Table 2] The Control Barrier Function filter is described as an 'optional' safety mechanism and does not appear in the configuration descriptions or the reported experiments. If it was not active in the 1,000-run or 25-run simulations, the paper should say so explicitly. As written, the abstract's mention of CBF filtering overstates what was actually evaluated, and the reader cannot tell whether the reported safety performance relied on the MPC constraints alone or on the CBF projection.
  5. [§IV.E.1 / §V.C / §IV.F] Equation (24) optimizes P* using the KSPDG seed distribution, and the Monte Carlo 'robustness' campaign in §V.C draws from the same distribution, as stated in §IV.F. The 94% success rate is therefore an in-sample estimate. To support the claimed robustness and to rule out that P* memorizes the seed distribution, the evaluation should include held-out initial conditions, a different adversarial scenario, or at minimum a cross-validated report of performance over seeds not used in the CEM optimization. The paper should also describe the seed sampling distribution and its relationship to the training distribution.
minor comments (6)
  1. [Eq. (8)] The input-rate constraint uses Δu_max but this quantity is not defined in the nomenclature or the text. Please define it.
  2. [§IV.E.2] The term 'capture lane is open' is used without definition. Clarify what geometric condition determines the capture lane and how it is computed.
  3. [Fig. 3 / Fig. 4] Panels (B)–(D) are referenced in the text, but the figure captions do not explicitly describe what each panel shows. Adding a clear per-panel description would improve readability.
  4. [§V.C] The paper says 'A total of 1000 closed-loop simulation runs were executed' but does not state the hardware, solver, or per-run computation time. Reporting the QP solver and average solve time would support the real-time feasibility claim.
  5. [Data Availability Statement] The statement that access 'may require institutional approval' is vague. If possible, provide a repository link or at least specify the conditions under which code and data can be shared.
  6. [References] Reference [28] is a 2017 arXiv preprint; consider citing the later peer-reviewed version if available. Also, the survey in reference [6] is listed without a publication venue; consider updating.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the adaptive-MPC claim rests on empirical comparison and explicit limitations, not on a self-defined or self-cited reduction.

full rationale

The paper's derivation chain is self-contained. The MPC cost and QP in Eqs. 15-16 are constructed from the discretized CW dynamics (Eq. 7) and standard prediction matrices (Eqs. 11-12); the predictive safety constraint (Eq. 20) is a first-order linearization of the KOZ distance condition, and the CBF filter (Eqs. 21-23) is a standard projection. The offline parameter optimizer (Eq. 24) fits P* by minimizing expected mission cost over the KSPDG seed distribution; this is a tuning step, not a definition of the evaluation statistic. The reported 94% and 72% success rates are measurements produced by running the closed-loop simulator, not quantities forced by Eq. 24. The paper explicitly disclaims formal recursive-feasibility and stability guarantees ('Formal recursive feasibility and closed-loop stability guarantees under arbitrary adaptive parameter updates remain outside the scope of the present work'), so no theorem is imported from author self-citation. There are no self-citations in the reference list, and the external citations (CW dynamics, MPC, CEM, CBF) are standard methods invoked for background, not load-bearing circular support. The adaptive-vs-fixed comparison may be open to an experimental-design critique — the Fixed Agent is only 'manually tuned' while the Adaptive Agent uses CEM-tuned P* — but unequal tuning effort is a validity confound, not a circular reduction of a claimed result to its inputs.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities; its burden is parameter fitting and modeling assumptions. There are at least six hand-set or fitted parameter groups, of which only P* is explicitly optimized (Eq. 24), and it is fitted to the same distribution later used for validation. The CW adequacy, fixed-opponent-policy, and transfer assumptions are domain assumptions the central empirical claims depend on. No independent falsifiable handle is provided for any tuning choice.

free parameters (6)
  • P* (offline-optimized base parameter vector: tracking weights, horizons, terminal multipliers, KOZ weights) = not disclosed
    Eq. 24 fits P* via CEM by minimizing expected KSPDG episode return; the same environment distribution is used for the headline 1000-run evaluation and the Fig. 5 comparison.
  • Phase-dependent tracking multipliers (4,25), (8,50), (12,80), (2.5,12) = as listed
    Section IV.E.2: hand-set multipliers applied to (w_L,pos, w_L,vel) at 500 m, 200 m, 80 m thresholds and final-approach condition.
  • Minimum-separation inflation constants 1.25 and 0.08 = 1.25, 0.08
    Eqs. 25-26: hand-chosen to shape the d_min,eff response to Bandit-Guard proximity; no sensitivity analysis provided.
  • KOZ relaxation levels (0.5x, 0.3x, 0) = 0.5, 0.3, 0 (multiples of wKOZ)
    Section IV.E.2: hand-chosen feasibility-recovery steps that disable or reduce KOZ weighting when the QP is infeasible.
  • CBF gains k0, k1 = not reported
    Section IV.D, Eq. 22: gains shaping the distance and closing-rate safety response are never given numerically.
  • Core MPC constants (gamma_K, lambda_s, term_base, term_final, Hp, Hu, dt, dKOZ, dmin, wKOZ) = not reported
    These constants define the QP that generates all reported trajectories; none are disclosed, preventing independent replication or analysis of constraint activation.
assumptions (5)
  • domain assumption CW linearized dynamics remain sufficiently accurate for the entire adversarial RPO interaction
    Section IV.A: the paper acknowledges the environment "may include nonlinear and dynamically evolving interactions" and asserts adequacy of the linearized model without quantifying CW model error against the truth model.
  • domain assumption Guard and Lady policies in KSPDG are representative adversarial behaviors and are independent of the Bandit controller
    Section V.A: Lady and Guard behavior is "governed by the challenge simulation framework" with undisclosed policies; success rates are only meaningful under this fixed-policy assumption.
  • domain assumption Optimizing P* against KSPDG returns (Eq. 24) transfers to the evaluation distribution
    The expectation in Eq. 24 is over the same initial-condition seed distribution used in the Monte Carlo evaluation; no held-out distribution is tested, so transfer is assumed rather than demonstrated.
  • domain assumption Slack-relaxed KOZ constraints still constitute a meaningful safety mechanism
    Section IV.C: slack variables permit KOZ violation when feasibility is at risk; safety claims rest on penalized slack use, not hard guarantees. The paper explicitly disclaims formal guarantees.
  • domain assumption Discrete-time CBF condition (Eq. 22) with estimated adversary acceleration provides valid safety filtering
    Section IV.D: the CBF guarantee from [24] assumes known dynamics; here the Guard acceleration is a PD-pursuit estimate, so the nominal guarantee does not strictly hold.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AI-Augmented Model Predictive Control for Safe and Adaptive Rendezvous and Proximity Operations." pith.science (2026). https://pith.science/paper/XZVPKX2R

@misc{pith2026260716630,
  author       = {Pith},
  title        = {Pith review of: AI-Augmented Model Predictive Control for Safe and Adaptive Rendezvous and Proximity Operations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XZVPKX2R}},
  note         = {Machine review of arXiv:2607.16630}
}
read the original abstract

Autonomous rendezvous and proximity operations (RPO) in adversarial orbital environments require guidance architectures balancing target pursuit, safety preservation, and real-time adaptability under dynamically evolving interaction conditions. Although learning-based approaches show promise, their application to safety-critical orbital robotics remains limited by concerns regarding interpretability, robustness, and constraint awareness. This work presents an adaptive Model Predictive Control (MPC) framework for autonomous spacecraft RPO in multi-agent adversarial scenarios. The proposed architecture combines a constrained receding-horizon MPC formulation with a data-driven supervisory tuning layer that adjusts controller parameters from offline closed-loop evaluation and online interaction geometry. Relative motion follows Clohessy-Wiltshire (CW) dynamics, enabling computationally efficient finite-horizon prediction and real-time quadratic optimization. The MPC formulation incorporates actuator limits, predictive keep-out-zone constraints, slack-variable feasibility handling, and optional Control Barrier Function (CBF) safety filtering. Rather than generating thrust commands directly, the adaptive layer modifies interpretable MPC parameters, including tracking weights, safety penalties, minimum-separation objectives, and keep-out-zone objectives. The framework was evaluated in the official Kerbal Space Program Differential Game (KSPDG) Capture-the-Satellite environment through Monte Carlo simulations. Results demonstrate improved closed-loop robustness, adaptive maneuvering behavior, and rendezvous performance compared with fixed-parameter MPC while preserving safety-aware operation and real-time feasibility, providing a modular, interpretable foundation for adaptive spacecraft RPO.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 6 canonical work pages

  1. [1]

    B., Mayne, D

    Rawlings, J. B., Mayne, D. Q., and Diehl, M. M.,Model Predictive Control: Theory, Computation, and Design, 2nd ed., Nob Hill Publishing, 2017

    2017

  2. [2]

    Constrainedmodelpredictivecontrol: Stabilityandoptimality,

    Mayne,D.,Rawlings,J.,Rao,C.,andScokaert,P.,“Constrainedmodelpredictivecontrol: Stabilityandoptimality,”Automatica, Vol. 36, No. 6, 2000, pp. 789–814. https://doi.org/10.1016/S0005-1098(99)00214-9

    doi:10.1016/s0005-1098(99)00214-9

  3. [3]

    F., and Bordons, C.,Model Predictive Control, Springer, 2007

    Camacho, E. F., and Bordons, C.,Model Predictive Control, Springer, 2007

    2007

  4. [4]

    (eds.),Modern Spacecraft Guidance, Navigation, and Control: From System Modeling to AI and Innovative Applications, Elsevier, 2022

    Pesce, V., Colagrossi, A., and Silvestrini, S. (eds.),Modern Spacecraft Guidance, Navigation, and Control: From System Modeling to AI and Innovative Applications, Elsevier, 2022. https://doi.org/10.1016/C2020-0-03563-2

    work page doi:10.1016/c2020-0-03563-2 2022

  5. [5]

    Fast, Safe, Propellant-Efficient Spacecraft Motion Planning Under Clohessy–Wiltshire–Hill Dynamics,

    Starek, J. A., Schmerling, E., Maher, G. D., Barbee, B. W., and Pavone, M., “Fast, Safe, Propellant-Efficient Spacecraft Motion Planning Under Clohessy–Wiltshire–Hill Dynamics,”Journal of Guidance, Control, and Dynamics, Vol. 40, No. 2, 2017, pp. 418–438. https://doi.org/10.2514/1.G001913

    work page doi:10.2514/1.g001913 2017

  6. [6]

    A Survey on Artificial Intelligence Trends in Spacecraft Guidance Dynamics and Control,

    Izzo, D., Märtens, M., and Pan, B., “A Survey on Artificial Intelligence Trends in Spacecraft Guidance Dynamics and Control,” , 2018. URL https://arxiv.org/abs/1812.02948

    arXiv 2018

  7. [7]

    SpaceGym: Discrete and Differential Games in Non-Cooperative Space Operations,

    Allen, R. E., Rachlin, Y., Ruprecht, J., Loughran, S., Varey, J., and Viggh, H., “SpaceGym: Discrete and Differential Games in Non-Cooperative Space Operations,”2023 IEEE Aerospace Conference, 2023, pp. 1–12. https://doi.org/10.1109/AERO55745. 2023.10115968

    arXiv 2023

  8. [8]

    A Model-Predictive Control Technique with Guaranteed Resolvability and Required Thruster Silent Times for Small-Body Proximity Operations,

    Carson, J. M., and Acikmese, B., “A Model-Predictive Control Technique with Guaranteed Resolvability and Required Thruster Silent Times for Small-Body Proximity Operations,”AIAA Guidance, Navigation, and Control Conference and Exhibit, 2006. https://doi.org/10.2514/6.2006-6780

    work page doi:10.2514/6.2006-6780 2006

Show all 29 references
  1. [9]

    D.,Orbital Mechanics for Engineering Students, 4th ed., Butterworth-Heinemann, Oxford, UK, 2019

    Curtis, H. D.,Orbital Mechanics for Engineering Students, 4th ed., Butterworth-Heinemann, Oxford, UK, 2019

    2019

  2. [10]

    Fehse, W.,Automated Rendezvous and Docking of Spacecraft, Cambridge Aerospace Series, Cambridge University Press, 2003. 32

    2003

  3. [11]

    Spacecraft Trajectory Planning with Avoidance Constraints Using Mixed-IntegerLinearProgramming,

    Richards, A., Feron, E., How, J., and Schouwenaars, T., “Spacecraft Trajectory Planning with Avoidance Constraints Using Mixed-IntegerLinearProgramming,”JournalofGuidanceControlandDynamics,Vol.25,2002. https://doi.org/10.2514/2.4943

    2002 doi

  4. [12]

    Safe Trajectories for Autonomous Rendezvous of Spacecraft,

    Breger, L., and How, J., “Safe Trajectories for Autonomous Rendezvous of Spacecraft,”Journal of Guidance Control and Dynamics, Vol. 31, 2008. https://doi.org/10.2514/1.29590

    2008 doi

  5. [13]

    Reinforcement-Learning-Enhanced Model Predictive Control with Application to Autonomous Planetary Landing,

    Federici, L., Benedikter, B., and Furfaro, R., “Reinforcement-Learning-Enhanced Model Predictive Control with Application to Autonomous Planetary Landing,”Journal of Guidance, Control, and Dynamics, Vol. 49, No. 3, 2026, pp. 788–805. https://doi.org/10.2514/1.G009534

    2026 doi

  6. [14]

    Reinforced Model Predictive Guidance and Control for Spacecraft Proximity Operations,

    Capra, L., Brandonisio, A., and Lavagna, M. R., “Reinforced Model Predictive Guidance and Control for Spacecraft Proximity Operations,”Aerospace, Vol. 12, No. 9, 2025. https://doi.org/10.3390/aerospace12090837

    2025 doi

  7. [15]

    Opponent Modeling in Deep Reinforcement Learning,

    He, H., Boyd-Graber, J. L., Kwok, K., and III, H. D., “Opponent Modeling in Deep Reinforcement Learning,”CoRR, 2016. URL https://arxiv.org/abs/1609.05559

    2016 arXiv

  8. [16]

    A Comprehensive Survey on Safe Reinforcement Learning,

    García, J., and Fernández, F., “A Comprehensive Survey on Safe Reinforcement Learning,”Journal of Machine Learning Research, Vol. 16, No. 42, 2015, pp. 1437–1480. URL http://jmlr.org/papers/v16/garcia15a.html

    2015

  9. [17]

    Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learning,

    Brunke, L., Greeff, M., Hall, A., Yuan, Z., Zhou, S., Panerati, J., and Schoellig, A., “Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learning,” 2021. https://doi.org/10.48550/arXiv.2108.06266

    2021 doi arXiv

  10. [18]

    Learning-Based Model Predictive Control: Toward Safe Learning in Control,

    Hewing, L., Wabersich, K., Menner, M., and Zeilinger, M., “Learning-Based Model Predictive Control: Toward Safe Learning in Control,”Annual Review of Control, Robotics, and Autonomous Systems, Vol. 3, 2020. https://doi.org/10.1146/annurev- control-090419-075625

    2020 doi

  11. [19]

    Learning-Based Model Predictive Control for Safe Exploration,

    Koller, T., Berkenkamp, F., Turchetta, M., and Krause, A., “Learning-Based Model Predictive Control for Safe Exploration,” 2018 IEEE Conference on Decision and Control (CDC), 2018, pp. 6059–6066. https://doi.org/10.1109/CDC.2018.8619572

    2018

  12. [20]

    Optimal and Autonomous Control Using Reinforcement Learning: A Survey,

    Kiumarsi, B., Vamvoudakis, K. G., Modares, H., and Lewis, F. L., “Optimal and Autonomous Control Using Reinforcement Learning: A Survey,”IEEE Transactions on Neural Networks and Learning Systems, Vol. 29, No. 6, 2018, pp. 2042–2062. https://doi.org/10.1109/TNNLS.2017.2773458

    2018

  13. [21]

    EfficientIterativeLinear-QuadraticApproximations forNonlinearMulti-PlayerGeneral-SumDifferentialGames,

    Fridovich-Keil,D.,Ratner,E.,Peters,L.,Dragan,A.D.,andTomlin,C.J.,“EfficientIterativeLinear-QuadraticApproximations forNonlinearMulti-PlayerGeneral-SumDifferentialGames,”2020IEEEInternationalConferenceonRoboticsandAutomation (ICRA), 2020, pp. 1475–1481. https://doi.org/10.1109/...

    2020

  14. [22]

    TerminalGuidanceSystemforSatelliteRendezvous,

    Clohessy,W.H.,andWiltshire,R.S.,“TerminalGuidanceSystemforSatelliteRendezvous,”JournaloftheAerospaceSciences, Vol. 27, No. 9, 1960, pp. 653–658. https://doi.org/10.2514/8.8704

    1960 doi

  15. [23]

    Soft Constraints and Exact Penalty Functions in Model Predictive Control,

    Kerrigan, E. C., and Maciejowski, J. M., “Soft Constraints and Exact Penalty Functions in Model Predictive Control,” Proceedings of the UKACC International Conference on Control, 2000. URL https://www-control.eng.cam.ac.uk/Homepage/ papers/cued_control_53.pdf. 33

    2000

  16. [24]

    Control Barrier Function Based Quadratic Programs for Safety Critical Systems,

    Ames, A. D., Xu, X., Grizzle, J. W., and Tabuada, P., “Control Barrier Function Based Quadratic Programs for Safety Critical Systems,”IEEE Transactions on Automatic Control, Vol. 62, No. 8, 2017, pp. 3861–3876. https://doi.org/10.1109/TAC.2016. 2638961

    2017 doi

  17. [25]

    ATutorialontheCross-EntropyMethod,

    deBoer,P.-T.,Kroese,D.P.,Mannor,S.,andRubinstein,R.Y.,“ATutorialontheCross-EntropyMethod,”AnnalsofOperations Research, Vol. 134, No. 1, 2005, pp. 19–67. https://doi.org/10.1007/s10479-005-5724-z

    2005 doi

  18. [26]

    Cross-Entropy Method Variants for Optimization,

    Moss, R. J., “Cross-Entropy Method Variants for Optimization,”CoRR, Vol. abs/2009.09043, 2020. https://doi.org/10.48550/ arXiv.2009.09043

    2009 doi arXiv

  19. [27]

    CompletelyDerandomizedSelf-AdaptationinEvolutionStrategies,

    Hansen,N.,andOstermeier,A.,“CompletelyDerandomizedSelf-AdaptationinEvolutionStrategies,”EvolutionaryComputation, Vol. 9, 2001, pp. 159–195. https://doi.org/10.1162/106365601750190398

    2001 doi

  20. [28]

    Towards A Rigorous Science of Interpretable Machine Learning,

    Doshi-Velez, F., and Kim, B., “Towards A Rigorous Science of Interpretable Machine Learning,”arXiv: Machine Learning,

  21. [2017]

    https://doi.org/10.48550/arXiv.1702.08608. 34

    doi arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.