Risk-aware planning must identify rare, costly outcomes using only a limited number of model evaluations. Monte Carlo rollouts spend this budget on noisy tail estimates, while planning from the mean can miss the tail entirely. We introduce quantile propagation planning (QPP), which propagates a small, fixed set of disturbance quantiles through the dynamics and integrates the resulting cost curve to evaluate both mean and tail risk. The method is exact at the propagated levels, converges as the quantile grid is refined, and admits a one-sided conservatism guarantee under stated structural conditions. In controlled overload and inventory tasks, QPP reduces realized tail cost and constraint violations relative to mean-rollout planning. On overload, it remains competitive with sampled evaluators that use substantially more trajectories. These results position QPP as a transparent, budget-controlled approach to tail-risk evaluation.
Introduction
Forecasts guide actions before their outcomes are known. Retail demand forecasts drive purchasing, inventory, logistics, and pricing ([10]); building and microgrid controllers schedule energy from uncertain load and weather ([9]); and portfolio optimizers are sensitive to errors in estimated returns and covariances ([15]). Learning-based model predictive control (MPC) faces the same issue: stochastic learned models must be used without losing constraint reliability ([13]). Optimization can therefore amplify a modest predictive error by selecting a different action and propagating its effect through repeated decisions.
Rare errors also have asymmetric consequences. A full-scale vehicle study reports that a deterministic MPC baseline using one dynamics model may lose control in adverse conditions, whereas a risk-averse controller accounts for uncertain friction and tire parameters ([17]). Worst-case protection, however, can make a plan too conservative to use. The practical goal is therefore to balance typical performance against costly tail outcomes, not merely to optimize either the mean or the worst case.
This balance becomes a statistical bottleneck when a planner ranks many candidates under a fixed model-call budget. At tail mass α, empirical CVaR uses only the worst α fraction of sampled trajectories; with α=0.1 and five rollouts, the estimate is controlled by the single most costly rollout. Concentration results formalize how the sample budget affects CVaR accuracy and selection error ([26]). Recent CVaR policy-optimization work likewise traces poor sample efficiency to underused non-tail trajectories ([20, 19]). Although those methods study policy learning, they expose the same scarcity of useful tail data.
Stochastic MPC represents uncertainty through distributions, scenarios, or probabilistic constraints, but the general nonlinear problem remains computationally demanding ([21]). Risk-aware planners use, for example, stochastic search over nonlinear beliefs ([34]) or sample-based CVaR constraints [17]. A recent nonlinear SMPC framework still reports an accuracy–computation trade-off among uncertainty-propagation methods [16]; analytic propagation avoids sampling variance but typically requires compatible distributional and value-function structure [30]. Mean-only planning is cheaper but can hide skew and harmful tails. The open question is how to rank many plans by mean and tail cost when each candidate can afford only a handful of model calls.
We introduce quantile propagation planning (QPP), which evaluates K fixed disturbance levels and integrates their piecewise-linear cost curve to obtain its mean and CVaR. QPP holds the same disturbance rank across time and coordinates, so it evaluates a comonotonic surrogate rather than the independently disturbed execution process. This shared-rank coupling can be conservative and may change which plan is preferred.
Our contributions are:
A deterministic level-curve evaluator that uses K model trajectories and does not require an ordering projection.
Coupled-process guarantees for exactness at the chosen levels, grid-refinement convergence, and one-sided conservatism under stated conditions.
An evaluation across controlled overload and inventory tasks and a fitted microgrid, with budget-matched sampling baselines and ablations that identify when QPP is effective.
Related Work
Sampling-based planning. PETS [6] propagates uncertainty through an ensemble. iCEM [25] improves CEM candidates, whereas the dsCEM-style (substituted point set) construction [33] replaces random candidates by deterministic samples. QPP instead changes candidate evaluation; the composition study therefore crosses QPP with both samplers and separates candidate-distribution effects from loop mechanics. On real-load data, the fixed-budget sampling comparison is a PETS-style bootstrap ensemble (§ 6.3).
Risk-aware objectives. The CVaR representation of Rockafellar and Uryasev [29] is what makes the level-curve tail analytically integrable. Risk-averse MPC supplies the dynamic-control context [31]. Stochastic-search CVaR control [34], sample-based CVaR constraints [17], quantile-gradient policy optimization [19], distributional RL [8 a], and epistemic value distributions [18] all retain sampled or learned distributions. QPP’s distinction is deterministic level-curve evaluation inside planning.
Deterministic uncertainty propagation. Non-Gaussian stochastic MPC has used deterministic quadrature with stochastic Galerkin [3], polynomial-chaos prediction and control [23], and monotonicity-based reachable sets [12]. QPP is narrower: it recursively propagates one scalar level through an evaluable one-step model and integrates a piecewise-linear mean/CVaR curve. We claim neither general accuracy nor computational dominance over those methods.
Learned distributional predictions. Local propagation of parametric laws [24], deep multi-step quantile MPC [4], and multi-horizon quantile forecasting (Wen et al., 2017) differ from QPP’s recursive reuse of one latent level in a one-step model. Distribution-guided diffusion planning [11], quantile-regularized representations [36], and point-estimate CEM [27] are related hosts or alternatives, not evidence for nonlinear learned-model performance here.
Figure 1. QPP in the closed-loop planning pipeline. Fixed disturbance levels generate coupled trajectories and an analytic mean–tail score; CEM ranks candidate plans, executes the first action, observes the next state, and replans.
Method
Setting. A world model is zt+1=f(zt,ut)+σ(zt,ut)⊙εt with state zt∈Rd, action ut, drift f, per-coordinate loading σ and a standardized scalar disturbance εt with quantile function F−1; a per-step cost c(zt,ut) is charged on the pre-step state. The planner chooses an open-loop u0:H−1 from z0 and executes it on the real, independently disturbed process; f, σ, c need only be evaluable in batch.
Quantile-vector state. Fix levels 0<q1<⋯<qK<1 (L={0.05,0.25,0.5,0.75,0.95} here). QPP replaces the S particles of a sampled rollout by K deterministic trajectories, one per level, all started at z0:
Trajectory k reuses qk at every step and coordinate. Thus the K states evaluate the comonotonic process Zt(U) (one uniform U shared across time and coordinates) at U=qk. Set Ct(k)=c(zt(k),ut); each step costs K model evaluations per candidate.
Level-curve scoring. At each step, let Ct(⋅) linearly interpolate the points (qk,Ct(k)). We extend it as a constant on [0,q1] and [qK,1], and let U∼Unif(0,1). Its mean is a trapezoid rule whose end weights include the boundary masses; omitting them creates bias that does not vanish as the interior grid is refined. We use α for the upper-tail probability mass, so CVaR0.1 is the mean of the worst 10%. We use the representation
CVaRα(X)=vmin{v+α−1E[(X−v)+]}
from Rockafellar and Uryasev [29], because (Ct−v)+ has an analytic integral on every linear segment. The implementation brackets v by the row minimum and maximum and uses 60 bisection steps; Appendix A.§ 2 gives the weights and cell integrals. The result is exact to floating-point precision for the constructed interpolant, not for the unobserved curve between or beyond the knots. The cost row need not be sorted: a non-monotone cost can make its largest value occur away from the largest qk, but integration still follows the level coordinate u.
Order restriction along K. Interpreting trajectory k as a marginal qk quantile requires each coordinate of Zt(U) to be non-decreasing in U (Proposition 1(c)). Pool-adjacent violators [2, 28] or rearrangement [5] can impose this order (Appendix A.§ 5), but QPP uses neither: level-curve scoring does not require ordered costs, and projecting a state would change the constructed trajectory whenever the condition fails (Appendix A.§ 4).
lower is better: a sum of per-step tail risks, a conservative surrogate for trajectory CVaR because subadditivity gives
t∑CVaRα[Ct]≥CVaRα[t∑Ct].
This surrogate uses the comonotonic coupling, not the independent process executed (§ 4, § 6). At λ=0 the score is risk-neutral for the coupled process, which differs from a mean rollout when c is nonlinear; at λ=1 it is the summed per-step tail. The level-curve evaluator is the central component, while λ is an optional decision preference. Appendix A.10 separately audits a single CVaR applied to the total trajectory cost.
Algorithm 1 collects the evaluator for one candidate; the planner evaluates candidates in batch. The same qk is retained throughout each trajectory, and the cost row is neither sorted nor projected.
Planner, baselines, cost. The outer loop is the cross-entropy method: draw P candidates from a diagonal Gaussian, clip, score, refit to the best E, and finally rescore the elite mean. Both the sampler (Gaussian, iCEM-style Pinneri et al. [25], or dsCEM-style Walker et al. [33]) and evaluator are modular. Monte Carlo uses S independent samples and averages the worst ⌈αS⌉; mean rollout propagates f alone; moment matching uses a full central-difference Jacobian with diagonal covariance (Appendix A.§ 1). QPP uses K model calls per candidate-step and O(KB) scoring for fixed B=60; the other costs are S calls plus a sort, one call, and 2d+1 calls, respectively. Thus S=K is the budget-matched QPP–Monte Carlo comparison, with Neval denoting the number of dynamics evaluations.
Algorithm 1: QPP level-curve evaluation for one candidate action sequence.
Input: State z0, actions u0:H−1, levels L, tail mass α, and weight λ
Input: Evaluable f, σ, and c, and disturbance quantile function F−1
Output: Score J and propagated states {zt(k)}
J←0 and z0(k)←z0 for k=1,…,K
fort=0,…,H−1do
A. Propagate the fixed levels
One uniform U is shared across steps and coordinates; Zt is the independent process, a fresh εt per step, on which plans are executed. The guarantees below concern this scalar ordered disturbance construction; they do not cover arbitrary multivariate or non-monotone learned dynamics.
The next result separates three claims: exactness at the selected knots, convergence as the grid is refined, and the stronger conditions under which the propagated levels are marginal quantiles. It applies to the states and level curves produced by Algorithm 1.
References
[1]David Angeli and Eduardo D. Sontag. Monotone control systems. IEEE Transactions on Automatic Control, 48(10):1684–1698, 2003. doi: 10.1109/tac.2003.817920.DOI
[2]Richard E. Barlow, D. J. Bartholomew, J. M. Bremner, and H. D. Brunk. Statistical Inference under Order Restrictions: The Theory and Application of Isotonic Regression. Wiley, London, 1972.
[3]Huishan Chen and Zheng Zhang. Stochastic model predictive control of autonomous systems with Non-Gaussian correlated uncertainty. arXiv preprint arXiv:1910.02994, 2019.arxiv.org/abs/1910.02994
[4]Yi-Ping Chen, Ying-Kuan Tsai, Vispi Karkaria, and Wei Chen. Uncertainty-aware digital twins: Robust model predictive control using time-series deep quantile learning. arXiv preprint arXiv:2501.10337, 2025.DOI
[5]Victor Chernozhukov, Iván Fernández-Val, and Alfred Galichon. Quantile and probability curves without crossing. Econometrica, 78(3):1093–1125, 2010. doi: 10.3982/ecta7880.DOI
[6]Kurtland Chua, Roberto Calandra, Rowan McAllister, and Sergey Levine. Deep reinforcement learning in a handful of trials using probabilistic dynamics models. In Advances in Neural Information Processing Systems (NeurIPS), 2018.arxiv.org/abs/1805.12114
[8]Will Dabney, Mark Rowland, Marc G. Bellemare, and Rémi Munos. Distributional reinforcement learning with quantile regression. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, volume 32(1), pages 2892–2901, 2018b. doi: 10.1609/aaai.v32i1.11791.DOI
[9]Ján Drgoňa, Javier Arroyo, Iago Cupeiro Figueroa, David Blum, Krzysztof Arendt, Donghun Kim, Enric Perarnau Ollé, Juraj Oravec, Michael Wetter, Draguna L. Vrabie, and Lieve Helsen. All you need to know about model predictive control for buildings. Annual Reviews in Control, 50: 190–232, 2020. doi: 10.1016/j.arcontrol.2020.09.001.DOI
[10]Robert Fildes, Shaohui Ma, and Stephan Kolassa. Retail forecasting: Research and practice. International Journal of Forecasting, 38(4):1283–1318, 2022. doi: 10.1016/j.ijforecast.2019.06.004.
[11]Shiqiang Gong. RS-Diffuser: Risk-sensitive diffusion planning with distributional value guidance. arXiv preprint arXiv:2606.27766, 2026.arxiv.org/abs/2606.27766
[12]Moritz Heinlein, Sankaranarayanan Subramanian, and Sergio Lucia. Robust model predictive control exploiting monotonicity properties. IEEE Transactions on Automatic Control, 70(9):6260–6267, 2025. doi: 10.1109/TAC.2025.3558137.DOI
[13]Lukas Hewing, Kim P. Wabersich, Marcel Menner, and Melanie N. Zeilinger. Learning-based model predictive control: Toward safe learning in control. Annual Review of Control, Robotics, and Autonomous Systems, 3:269–296, 2020. doi: 10.1146/annurev-control-090419-075625.DOI
[14]Rob Kaas, Jan Dhaene, and Marc J. Goovaerts. Upper and lower bounds for sums of random variables. Insurance: Mathematics and Economics, 27(2):151–168, 2000. doi: 10.1016/s0167-6687(00)00060-3.DOI
[15]Petter N. Kolm, Reha Tütücü, and Frank J. Fabozzi. 60 years of portfolio optimization: Practical challenges and current trends. European Journal of Operational Research, 234(2):356–371, 2014. doi: 10.1016/j.ejor.2013.10.060.DOI
[16]Daniel Landgraf, Andreas Völz, and Knut Graichen. A software framework for stochastic model predictive control of nonlinear continuous-time systems (GRAMPC-S). Optimization and Engineering, 27:193–228, 2026. doi: 10.1007/s11081-025-10006-z.DOI
[17]Thomas Lew, Marcus Greiff, Franck Djemou, Makoto Suminaka, Michael Thompson, and John Subosits. Risk-averse model predictive control for racing in adverse conditions. arXiv preprint arXiv:2410.17183, 2024.arxiv.org/abs/2410.17183
[18]Carlos E. Luis, Alessandro G. Bottero, Julia Vinogradska, Felix Berkenkamp, and Jan Peters. Value-distributional model-based reinforcement learning. arXiv preprint arXiv:2308.06590, 2023.arxiv.org/abs/2308.06590
[19]Yudong Luo and Erick Delage. Boosting CVaR policy optimization with quantile gradients. In Proceedings of the 43rd International Conference on Machine Learning, volume 306 of Proceedings of Machine Learning Research, pages 82196–82219. PMLR, 2026. URL https://proceedings.mlr.press/v306/luo26b.html.arxiv.org/abs/2601.22100
[20]Harry Mead, Clarissa Costen, Bruno Lacerda, and Nick Hawes. Return capping: Sample-efficient CVaR policy gradient optimisation. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 43503–43518. PMLR, 2025. URL https://proceedings.mlr.press/v267/mead25a.html.arxiv.org/abs/2504.20887
[21]Ali Mesbah. Stochastic model predictive control: An overview and perspectives for future research. IEEE Control Systems Magazine, 36(6):30–44, 2016. doi: 10.1109/MCS.2016.2602087.DOI
[22]Clayton Miller, Anjukan Kathirgamanathan, Bianca Picchetti, Pandarasamy Arjunan, June Young Park, Zoltan Nagy, Paul Raftery, Brodie W. Hobson, Zixiao Shi, and Forrest Meggers. The building data genome project 2, energy meter data from the ASHRAE great energy predictor III competition. Scientific Data, 7:368, 2020. doi: 10.1038/s41597-020-00712-x.DOI
[23]Guanru Pan, Ruchuan Ou, and Timm Faulwasser. On a stochastic fundamental lemma and its use for data-driven optimal control. IEEE Transactions on Automatic Control, 68(10):5922–5937, 2023. doi: 10.1109/TAC.2022.3232442.DOI
[24]Felix Petersen, Aashwin Mishra, Hilde Kuehne, Christian Borgelt, Oliver Deussen, and Mikhail Yurochkin. Uncertainty quantification via stable distribution propagation. In International Conference on Learning Representations (ICLR), 2024.arxiv.org/abs/2402.08324
[25]Cristina Pinneri, Shambhuraj Sawant, Sebastian Blaes, Jan Achterhold, Joerg Stueckler, Michal Rolinek, and Georg Martius. Sample-efficient cross-entropy method for real-time planning. In Proceedings of the 2020 Conference on Robot Learning, volume 155 of Proceedings of Machine Learning Research, pages 1049–1065. PMLR, 2021. URL https://proceedings.mlr.press/v155/pinneri21a.html.arxiv.org/abs/2008.06389
[26]L. A. Prashanth, Krishna Jagannathan, and Ravi Kolla. Concentration bounds for CVaR estimation: The cases of light-tailed and heavy-tailed distributions. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 5577–5586. PMLR, 2020. URL https://proceedings.mlr.press/v119/1-a-20a.html.
[27]Farhad Rezazadeh, Amir Ashtari Gargari, Hatim Chergui, Sandra Lagen, Merouane Debbah, Houbing Song, and Lingjia Liu. Agentic world modeling for 6G: Near-real-time generative state-space reasoning. arXiv preprint arXiv:2511.02748, 2025.arxiv.org/abs/2511.02748
[28]Tim Robertson, F. T. Wright, and R. L. Dykstra. Order Restricted Statistical Inference. Wiley, Chichester, 1988.
[29]R. Tyrrell Rockafellar and Stanislav Uryasev. Optimization of conditional value-at-risk. The Journal of Risk, 2(3):21–41, 2000. doi: 10.21314/jor.2000.038.DOI
[30]Shishir Sharma and Doina Precup. Analytic planning under uncertainty with moment closure. In Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence, volume 337 of Proceedings of Machine Learning Research, pages 6165–6177. PMLR, 2026. URL https://proceedings.mlr.press/v337/sharma26a.html.arxiv.org/abs/2608.02519
[31]Pantelis Sopasakis, Domagoj Herceg, Alberto Bemporad, and Panagiotis Patrinos. Risk-averse model predictive control. Automatica, 100:281–288, 2019. doi: 10.1016/j.automatica.2018.11.022.DOI
[33]Markus Walker, Daniel Frisch, and Uwe D. Hanebeck. Sample-efficient and smooth cross-entropy method model predictive control using deterministic samples. In Proceedings of the 2026 American Control Conference (ACC), pages 4629–4636, New Orleans, Louisiana, 2026. IEEE. ISBN 979-8-3315-9382-7.arxiv.org/abs/2510.05706
[34]Ziyi Wang, Oswin So, Keuntaek Lee, and Evangelos A. Theodorou. Adaptive risk sensitive model predictive control with stochastic search. In Proceedings of the 3rd Conference on Learning for Dynamics and Control, volume 144 of Proceedings of Machine Learning Research, pages 510–522. PMLR, 2021. URL https://proceedings.mlr.press/v144/wang21b.html.arxiv.org/abs/2009.01090
[36]Zhoushun Yu, Xiaoyu Hu, and Xiangyu Xu. QOWorld: Quantile-quantile matching for world model regularization. arXiv preprint arXiv:2607.28415, 2026.arxiv.org/abs/2607.28415