Risk-Aware Model Predictive Control

• Risk-aware MPC is a receding-horizon control framework that incorporates risk measures to balance nominal performance with mitigation of rare, severe events.
• It employs diverse formulations including CVaR, chance constraints, and learned uncertainty models to address stochastic dynamics and ambiguity.
• The approach has practical applications in autonomous driving, robot navigation, and microgrid control, with some methods offering formal stability guarantees.

Searching arXiv for recent and foundational papers on risk-aware MPC to support the article. arXiv search: query="risk-aware model predictive control", max_results=10, sort_by="relevance" Risk-aware Model Predictive Control (MPC) denotes a family of receding-horizon control methods in which future control actions are chosen not only by nominal performance criteria but also by explicit representations of uncertainty, tail events, or safety violations. In the recent literature, this family includes time-consistent dynamic risk metrics for stochastic control, chance-constrained and CVaR-constrained MPC, distributionally robust formulations over ambiguity sets, and learned risk surrogates based on predictive world models or data-driven uncertainty quantification (Chow et al., 2015, Sun et al., 26 Feb 2026). Across these variants, the common objective is not merely to predict future trajectories, but to rank or constrain them according to a notion of risk that is compatible with the control task, the uncertainty model, and the desired balance between performance and safety.

1. Foundations and scope

The canonical MPC template optimizes a finite-horizon input sequence subject to dynamics and constraints,

$$\min_{\mathbf{a}_{t:t+H-1}} \sum_{k=0}^{H-1} \ell(x_{t+k}, a_{t+k}) \quad \text{s.t. } x_{t+k+1} = f(x_{t+k}, a_{t+k}), \; (x,a) \in \mathcal{X}\times\mathcal{U}.$$

Risk-aware MPC modifies this template by replacing purely nominal objectives or hard worst-case constraints with risk-sensitive objectives, probabilistic constraints, coherent risk measures, or explicit risk surrogates (Sun et al., 26 Feb 2026, Chow et al., 2015).

A central conceptual distinction in the literature is between robust, stochastic, and risk-averse formulations. Conventional robust MPC enforces constraints for all uncertainty realizations in a prescribed set, which yields deterministic guarantees but can be overly conservative. Stochastic MPC instead enforces probabilistic constraints, typically of the form

$$\mathbb{P}\!\big(g(x_{t:t+H}, a_{t:t+H-1}) \le 0\big) \ge 1-\epsilon.$$

Risk-averse MPC occupies an intermediate and more general position: it can recover expectation-based, chance-constrained, worst-case, or ambiguity-aware formulations depending on the chosen risk measure and uncertainty model (Li, 15 Jul 2025, Schießl et al., 13 Apr 2026).

Several papers make this unification explicit. For linear systems with multiplicative uncertainty, time-consistent dynamic risk metrics yield a framework that spans risk-neutral and worst-case MPC through polytopic risk envelopes (Chow et al., 2015). For constrained nonlinear Markovian switching systems, coherent Markov risk measures similarly interpolate between stochastic and robust control while preserving dynamic programming structure (Sopasakis et al., 2017). In microgrid operation, nested AVaR on a scenario tree is used precisely to interpolate between risk-neutral stochastic MPC at $\alpha=1$ and worst-case robust MPC at $\alpha=0$ (Hans et al., 2018).

2. Risk representations and objective structures

A common misconception is that risk-aware MPC is synonymous with chance constraints. The literature is broader. Some formulations constrain violation probabilities, some constrain tail severity, some optimize coherent risk measures of accumulated cost, and some use learned surrogates for hazardous outcomes.

In stochastic MPC with risk-averse constraints, a standard distinction is between Value-at-Risk and Conditional Value-at-Risk. For a real-valued random variable $Z$ and $\alpha\in(0,1)$,

$$\operatorname{VaR}_{1-\alpha}(Z):=\inf\{t\in\mathbb{R}:\ \mathbb{P}(Z\le t)\ge 1-\alpha\},$$

and

$$\operatorname{CVaR}_{\alpha}(Z)=\min_{t\in\mathbb{R}}\left\{t+\frac{1}{1-\alpha}\mathbb{E}\big[(Z-t)_+\big]\right\}.$$

The key motivation is that chance constraints control only violation frequency, whereas CVaR-type constraints also account for the magnitude of rare but severe violations (Schießl et al., 13 Apr 2026).

More general coherent and law-invariant risk measures also appear. Spectral risk measures are represented as

$\rho(Z)=\int_0^1 \phi(v)\,F_Z^{-1}(v)\,dv,$

with a nondecreasing weighting function $\phi$, which allows direct control of tail emphasis without committing to a single parametric form such as CVaR (Eom et al., 2 Jun 2026). In time-consistent risk-averse MPC, coherent one-step risk mappings are nested across the horizon, which is essential for Bellman-style dynamic programming and for Lyapunov-type stability arguments (Chow et al., 2015, Sopasakis et al., 2017).

Not all risk-aware MPC schemes use formal coherent risk measures. In autonomous driving, RaWMPC defines horizon risk by an explicit rollout cost that combines progress and predicted event probabilities,

$$\mathbb{P}\!\big(g(x_{t:t+H}, a_{t:t+H-1}) \le 0\big) \ge 1-\epsilon.$$0

with event-type weights $$\mathbb{P}\!\big(g(x_{t:t+H}, a_{t:t+H-1}) \le 0\big) \ge 1-\epsilon.$$1 and discount $$\mathbb{P}\!\big(g(x_{t:t+H}, a_{t:t+H-1}) \le 0\big) \ge 1-\epsilon.$$2 (Sun et al., 26 Feb 2026). In that formulation, the paper explicitly states that chance constraints and CVaR are not used. Related trajectory planners also embed risk as soft artificial potentials for infrastructure and objects, or as normalized conflict-risk fields derived from predictive distributions of surrounding vehicles (Ploeg et al., 2022, Huang et al., 2023).

3. Uncertainty modeling and learning

The effectiveness of risk-aware MPC depends on how uncertainty is represented. The recent literature shows a marked shift from fixed analytic uncertainty sets toward learned predictive models and data-adaptive ambiguity sets.

One line of work learns predictive distributions of exogenous agents. A Bayesian LSTM-based vehicle planner predicts mean and variance of surrounding vehicles’ future positions via Monte Carlo dropout, parameterizes a bivariate Normal distribution for each vehicle, and converts those distributions into normalized spatial risk fields queried by the MPC over a horizon of $$\mathbb{P}\!\big(g(x_{t:t+H}, a_{t:t+H-1}) \le 0\big) \ge 1-\epsilon.$$3 seconds (Huang et al., 2023). In crowd navigation, moment-based ambiguity sets are constructed from Trajectron++ rollouts by estimating pedestrian mean and covariance, and the control problem uses a distributionally robust CVaR relaxation of collision-probability chance constraints (Ryu et al., 2024).

A second line learns model error or uncertainty sets directly from operational data. RAAR-MPC introduces a medium-frequency risk assessment engine based on Gaussian process regression and active learning to construct a Learned Prediction-Error Set, then combines it with an adaptive safety margin in a dual-timescale robust MPC architecture (Li, 15 Jul 2025). Online learning for linear uncertain systems uses Dirichlet process mixture models to infer multimodal disturbance structure and constructs ambiguity sets from componentwise moment information, which are then used in distributionally robust CVaR-tightened MPC (Ning et al., 2020). For Markovian switching systems with unknown transition probabilities, ambiguity sets around empirically estimated transition rows are updated online and used inside a risk-averse optimal control problem that remains less conservative than fully robust MPC (Schuurmans et al., 2020, Schuurmans et al., 2020).

A third line abandons parametric distribution modeling altogether. Distribution-free conformal spectral risk control builds calibrated prediction sets guaranteeing that spectral risk remains below a user-specified threshold without assuming a parametric uncertainty distribution (Eom et al., 2 Jun 2026). This is then converted into tightened deterministic MPC constraints through a Lipschitz bound on the safety function.

World-model-based methods extend uncertainty modeling into high-dimensional perception-action loops. RaWMPC predicts future state tokens, semantic segmentations, event probabilities, and ego states under candidate action sequences, then uses those predicted consequences for risk evaluation (Sun et al., 26 Feb 2026). MonoMPC instead learns an action-conditioned Gaussian distribution over worst-case obstacle clearance from monocular perception and uses an MMD-based surrogate for collision risk inside sampling-based MPC (Sharma et al., 10 Aug 2025).

4. Receding-horizon architectures and numerical solution

The control architectures used in risk-aware MPC are highly heterogeneous. Some retain the classical convex-QP or NLP structure; others rely on sampling, tree search, or hybrid optimization over scenario trees.

In linear stochastic MPC with risk-averse constraints, the indirect feedback approach fixes a stabilizing linear feedback $$\mathbb{P}\!\big(g(x_{t:t+H}, a_{t:t+H-1}) \le 0\big) \ge 1-\epsilon.$$4 and parameterizes the control as

$$\mathbb{P}\!\big(g(x_{t:t+H}, a_{t:t+H-1}) \le 0\big) \ge 1-\epsilon.$$5

This separates nominal prediction from stochastic error dynamics and yields precomputable risk margins, so the online MPC remains a convex program with tightened linear constraints (Schießl et al., 13 Apr 2026). Tube-based robust MPC with time-varying tightening appears in RAAR-MPC, where the inner loop solves a convex QP while a learned uncertainty set and adaptive margin are updated on slower timescales (Li, 15 Jul 2025).

Nonlinear motion planning methods often solve direct optimal control problems. A long-horizon trajectory generator for automated driving uses a nonlinear kinematic bicycle model with Gaussian artificial potential fields for infrastructure and objects, and solves the resulting NLP with CasADi and IPOPT at every sampling time (Ploeg et al., 2022). Racing in adverse conditions uses a sample-based approximation of a CVaR-constrained stochastic optimal control problem, closed-loop scenario particles, and sequential quadratic programming with GPU parallelization to handle nonlinear tire-force uncertainty in real time (Lew et al., 2024).

Sampling-based architectures are increasingly common when perception and learned predictors are prominent. RaWMPC performs sampling-based MPC over candidate action sequences, evaluates each by world-model rollout, and selects the minimum-cost sequence without gradient-based refinement (Sun et al., 26 Feb 2026). MonoMPC uses a random-shooting/CEM-style optimizer over control sequences, evaluates each using a learned clearance distribution and empirical MMD risk, and updates the sampling distribution from elite candidates (Sharma et al., 10 Aug 2025). In crowd navigation, constrained CEM with GPU parallelization is used to enforce distributionally robust CVaR collision constraints in real time (Ryu et al., 2024).

Hybrid and symbolic variants also exist. For vehicle motion planning with discrete motion primitives, a depth-first search over a hybrid automaton computes cumulative risk over a receding horizon (Huang et al., 2023). For runtime temporal logic tasks, mixed-integer formulations embed signal temporal logic, slack-based probabilistic reachable tube bounds, and accept/reject decisions for newly arriving specifications (Engelaar et al., 2024). For priced timed automata, first-order logic encoding and Z3 support path planning with a scalarized cost-risk objective under automaton updates driven by failures (Anbarani et al., 2022).

5. Representative application domains

Autonomous driving is one of the most active application areas. RaWMPC addresses end-to-end autonomous driving without expert action supervision by predicting candidate outcomes with a world model and selecting low-risk actions through explicit event-based cost evaluation; it reports superior performance in both in-distribution and out-of-distribution scenarios and emphasizes decision interpretability through predicted semantics and event probabilities (Sun et al., 26 Feb 2026). Learned predictive conflict-risk fields have also been used for probabilistic vehicle motion planning in traffic (Huang et al., 2023), while long-horizon risk-averse motion planning with artificial potentials has been demonstrated in highway and urban scenarios, including proactive lane changes and braking for crossing road users (Ploeg et al., 2022). At the limits of handling, CVaR-constrained scenario MPC has been used to maintain reliable racing performance under uncertain friction and tire parameters (Lew et al., 2024).

Robot navigation provides a complementary set of formulations. Distributionally robust chance-constrained MPC in crowds uses moment ambiguity sets and CVaR relaxations to guarantee collision-probability bounds under uncertain pedestrian predictions (Ryu et al., 2024). Distribution-free conformal spectral risk control converts calibrated prediction sets into deterministic margins inside MPC, improving safety and solve time relative to an SAA-based baseline in obstacle avoidance (Eom et al., 2 Jun 2026). Runtime temporal-logic MPC extends risk-aware control to dynamically assigned specifications under unbounded disturbances (Engelaar et al., 2024). Monocular navigation in cluttered environments uses learned clearance distributions and an MMD surrogate for collision risk rather than explicit geometric depth-based checking (Sharma et al., 10 Aug 2025).

Risk-aware MPC also appears outside mobile robotics. In islanded microgrids, nested AVaR over a scenario tree yields a mixed-integer quadratically constrained quadratic program that trades operating cost against resilience to renewable and load uncertainty (Hans et al., 2018). In cyber-physical scheduling and manufacturing, priced timed automata MPC uses a risk-weighted path objective and topology updates to respond to failures (Anbarani et al., 2022). Iterative distributionally robust MPC further shows how risk constraints can be tightened progressively as data accumulates across repeated tasks (Zolanvari et al., 2023).

6. Guarantees, misconceptions, and limitations

A major dividing line in the literature is between empirical risk reduction and formal closed-loop guarantees. Several frameworks provide recursive feasibility and stability theorems. Linear stochastic MPC with risk-averse constraints establishes recursive feasibility, closed-loop satisfaction of law-invariant risk constraints, and averaged closed-loop near-optimality via stochastic dissipativity (Schießl et al., 13 Apr 2026). Time-consistent risk-averse MPC for multiplicative uncertainty and Markovian switching systems provides Lyapunov-type stability conditions and stabilizing terminal constructions (Chow et al., 2015, Sopasakis et al., 2017). Runtime temporal-logic MPC proves recursive feasibility and probabilistic satisfaction of accepted specifications (Engelaar et al., 2024). RAAR-MPC proves recursive feasibility and convergence of empirical risk to a user-defined level under its dual-timescale adaptation (Li, 15 Jul 2025).

Other methods are explicit about the absence of such guarantees. RaWMPC states that safety gains are empirical and that robustness is operationalized through predicted violation probabilities rather than chance constraints or CVaR (Sun et al., 26 Feb 2026). MonoMPC likewise reports no formal probabilistic safety guarantees, since its MMD surrogate encourages residual distributions toward feasibility but does not certify a chance constraint (Sharma et al., 10 Aug 2025). Potential-field motion planners assume exact or deterministic predictions of other objects and leave formal uncertainty treatment to future work (Ploeg et al., 2022).

Another common misconception is that more risk aversion is always preferable. Across the literature, stronger risk aversion generally improves safety margins or tail performance, but it can increase conservatism, cost, or computational burden. Microgrid control reports the best overall trade-off at an intermediate AVaR level rather than at either extreme (Hans et al., 2018). Crowd navigation shows that smaller allowable collision probability increases minimum robot-human distance but raises positional cost (Ryu et al., 2024). In racing, higher conservatism improves reliability under wet patches but sacrifices peak performance (Lew et al., 2024).

Current limitations recur across domains. Scenario-tree and sample-based methods scale poorly with horizon and uncertainty dimension. Learning-based methods depend on the calibration of predictive uncertainty, the representativeness of operational data, and the treatment of distribution shift. Distribution-free methods avoid parametric misspecification but typically provide marginal rather than conditional guarantees (Eom et al., 2 Jun 2026). This suggests that risk-aware MPC is best understood not as a single algorithmic recipe, but as a design space in which the choice of risk measure, uncertainty representation, and optimization architecture must be matched to the application’s model fidelity, computational budget, and required guarantee level.