Distributionally-Robust Nonlinear Steering

• Distributionally-robust nonlinear steering is a method for controlling nonlinear systems under uncertainty using ambiguity sets to safeguard performance and constraints.
• Key approaches include tube-based DR-MPC, feedback linearization, and convex reformulations via Wasserstein, MMD, and moment-based ambiguity sets.
• Empirical studies in robotics and process control demonstrate its capability to maintain safety and enhance performance even under significant model uncertainties.

Distributionally-robust nonlinear steering refers to the design of control or planning strategies for nonlinear systems under uncertainty, where performance and constraints must be respected for all probability distributions within a specified ambiguity set. These sets model epistemic and aleatoric uncertainty in model parameters, dynamics, or disturbance distributions. The approach encompasses robust model predictive control, nonlinear motion planning, and risk-averse optimal control, with distributional ambiguity typically quantified using metrics such as the Wasserstein distance or kernel-based divergences. Methods in this class leverage convex reformulations, scenario-based tightening, and feedback linearization to maintain tractability even for general nonlinear and nonconvex constraints.

1. Foundational Problem Setting

Distributionally-robust nonlinear steering generally addresses discrete-time nonlinear stochastic dynamical systems

$x_{k+1} = f(x_k, u_k) + w_k,$

where $x_k \in \mathbb{R}^n$ is the state, $u_k \in \mathbb{R}^m$ is the control input, and $w_k$ is an exogenous disturbance with unknown nonparametric distribution $P_0$ (or satisfying only moment or support constraints). The primary objective consists of finding a control sequence or feedback policy that ensures state and input constraints—and typically chance or risk constraints—are satisfied for all $P$ belonging to an ambiguity set $\mathcal{P}$ centered around empirically observed or nominal distributions.

A canonical optimization-based nonlinear steering formulation in this context is

$$\begin{aligned} \min_{u_{0:H-1}} &\quad \ell(x_{0:H}, u_{0:H-1}) \ \text{s.t.} &\quad x_{k+1} = f(x_k, u_k) + w_k, \quad k=0,\ldots,H-1 \ &\quad \sup_{P \in \mathcal{P}} P\big[h(x_k, u_k, w_k) > 0\big] \leq \alpha, \ \forall k, \end{aligned}$$

where $h(x_k, u_k, w_k)$ is a possibly nonconvex constraint function and $\alpha$ is the permitted violation probability.

The outstanding challenge is that the true law of the disturbances or process deviations is not known, only sampled or estimated from data, or modeled to belong to a set of plausible distributions, thus leading to the need for robustification against a family of distributions.

2. Distributional Ambiguity Sets

Specification and efficient use of ambiguity sets are central.

• Wasserstein Ambiguity Sets: The most prominent approach in recent literature is to define the ambiguity set $x_k \in \mathbb{R}^n$0 as a ball around an empirical or nominal distribution $x_k \in \mathbb{R}^n$1 in the 1-Wasserstein or 2-Wasserstein metric,

$x_k \in \mathbb{R}^n$2

where $x_k \in \mathbb{R}^n$3 is the $x_k \in \mathbb{R}^n$4-Wasserstein distance and $x_k \in \mathbb{R}^n$5 is a radius tuned using statistical bounds or bootstrap methods. Wasserstein balls have the unique property of capturing both the spread and location of distributions and allow data-driven uncertainty modeling with finite-sample guarantees (Zhong et al., 2023, Gahlawat et al., 4 Sep 2025, Zhong et al., 2022).

• Moment-Based Ambiguity Sets: An alternative is to use sets defined by mean and covariance, leading to tractable conic or SDP reformulations for certain constraints (see (Safaoui et al., 2021, Renganathan et al., 2022)). For instance,

$x_k \in \mathbb{R}^n$6

• Kernel-based/MMD Ambiguity Sets: Maximum Mean Discrepancy (MMD) balls in a reproducing kernel Hilbert space offer a dimension-independent, nonparametric alternative, with ambiguity sets of the form

$x_k \in \mathbb{R}^n$7

where $x_k \in \mathbb{R}^n$8 is the feature map associated with the chosen kernel (Nemmour et al., 2022, Nemmour et al., 2021).

• Gradient-Regularized and Scenario-Based Sets: Distributional robustness can also be approximated through gradient-norm or scenario-based regularization, which penalizes the effect of distributional shift in the constraint functions (Nemmour et al., 2021).

3. Core Algorithmic Approaches

Tube-based and Feedback Linearization Approaches

• Tube-based DR-MPC: These methods construct a “tube” around the nominal trajectory to bound reachable states under worst-case disturbances from the ambiguity set. At each MPC step, the nominal nonlinear model is linearized, and a robust tube is computed through convex QPs or SOCPs, often leveraging duality for finite-sample Wasserstein or MMD balls (Zhong et al., 2023, Zhong et al., 2022, Nemmour et al., 2022). The control optimization problem is constructed to enforce tightened constraints, accounting for the maximum deviation permitted by the ambiguity set.
• Iterative Linear Quadratic Regulator (iLQR): Local iterative LQR solves are embedded within the DRNMPC, where at each iteration the nonlinear system is linearized along the nominal trajectory, and a feedback policy is computed to minimize cost and control ambiguity propagation. Tightened constraints are derived from the dual representation of the worst-case expectation over the ambiguity set (cf. (Zhong et al., 2023)).

Representative Algorithm Skeleton

Step	Description	Reference
Linearize dynamics	about the current nominal trajectory	(Zhong et al., 2023)
Solve feedback policy	using Riccati recursion or convex QP (iLQR)	(Zhong et al., 2023)
Compute tube/back-off	propagate Wasserstein/MMD ambiguity via duality or explicit recursion	(Zhong et al., 2023, Nemmour et al., 2022)
Solve tightened MPC	optimize nominal trajectory under tightened constraints	(Zhong et al., 2022, Zhong et al., 2023)
Apply first control input	recede horizon; update with new state	all

Dual and Convex Reformulations

• Wasserstein Duality: The inner supremum over the ambiguity set is converted into a convex regularization term, often resulting in additional tightening of constraints proportional to the ambiguity radius $x_k \in \mathbb{R}^n$9 and the local Lipschitz modulus or gradient norm of the constraint function (Nemmour et al., 2021, Zhong et al., 2023).
• MMD Kernel Duality and CVaR Approximation: Through strong duality in the RKHS, the distributional constraint is equivalently imposed as a regularized empirical risk, often approximated via the CVaR of the constraint function under the empirical law (Nemmour et al., 2022).
• Scenario-based Regularization: The empirical distribution is used directly, with constraints robustified via Lipschitz or gradient-norm penalties to cover possible distributional shifts (Nemmour et al., 2021). This approach is computationally light but less tight than full Wasserstein or MMD approaches.

4. Risk Constraints and Tractable Safe Sets

Distributionally robust chance constraints are at the heart of safety in nonlinear steering under uncertainty.

• Worst-case Probability/Chance Constraints: For half-space constraints $u_k \in \mathbb{R}^m$0, the worst-case violation probability under a moment-based ambiguity set admits closed-form expressions, e.g.,

$u_k \in \mathbb{R}^m$1

which can be enforced by a deterministic linear-conic tightening (Safaoui et al., 2021, Renganathan et al., 2022).

• Distributionally Robust CVaR Constraints: For general nonlinear constraint functions, the distributionally robust CVaR over a Wasserstein ball is used:

$u_k \in \mathbb{R}^m$2

yielding tractable reformulations for polytopic and quadratic constraints (Gahlawat et al., 4 Sep 2025).

• Ambiguity Propagation and Reachable Sets: The propagation of ambiguity through the nonlinear dynamics forms tubes of growing radius, explicitly characterized in terms of the system’s closed-loop gain and the Wasserstein radius (Zhong et al., 2023).

5. Examples and Empirical Verification

Multiple empirical studies demonstrate the efficacy of distributionally-robust nonlinear steering:

• Nonlinear Mass-Spring System and CSTR: Tube-based DR-MPC with Wasserstein ambiguity was empirically tested on these systems, revealing robust constraint satisfaction and improved performance over stochastic MPC under significant plant-model mismatch (Zhong et al., 2022).
• Double Integrator with Nonconvex Constraints: MMD-DRCCP-enabled steering in tube-based MPC achieves constraint violation rates matching target levels (e.g., empirical violation below $u_k \in \mathbb{R}^m$3 for 30 simulated trajectories), with bootstrap-based ambiguity radii (Nemmour et al., 2022).
• Risk-Aware Motion Planning for Robots: High-level reference trajectories from a distributionally robust RRT* (RANS-RRT*) satisfy robustified chance constraints, and subsequent tracking via robust LQR or NMPC ensures safety under heavy-tailed noise models (Safaoui et al., 2021, Renganathan et al., 2022).
• Covariance Steering under Model Uncertainty: An L₁-adaptive law yields an online bound (\emph{certificate}) on the Wasserstein distance between the true and nominal state; high-level covariance-steering planning then enforces state and safety constraints via DR-CVaR tightening (Gahlawat et al., 4 Sep 2025).

6. Advantages, Limitations, and Future Research

Distributionally-robust nonlinear steering offers enhanced out-of-sample and finite-sample safety guarantees compared to classical stochastic MPC and scenario-based MPC, particularly when the true disturbance distribution is unknown or changed from that assumed during controller design. Key points:

Advantages

• Finite-sample coverage and quantifiable robustness to distributional shift (Nemmour et al., 2022, Zhong et al., 2023, Nemmour et al., 2021).
• Applicability to general nonlinear, nonconvex constraints via MMD kernels or gradient-norm regularization.
• Convex reformulations and efficient solvers for tube-based, constraint-tightened MPC.
• Data-driven ambiguity tuning, with bootstrap schemes for kernel/MMD sets.
• Predictable and safe performance when combined with robust adaptive elements (e.g., L₁-adaptive control for covariance steering (Gahlawat et al., 4 Sep 2025)).

Limitations

• Local linearization restricts guarantee strength to small error/tube regimes in strongly nonlinear systems (Zhong et al., 2023, Zhong et al., 2022).
• Solving a sequence of nonlinear or conic programs online increases computational burden, especially with inner maximization over distributions.
• Scenario or sample complexity may become high for stringent risk levels.
• Wasserstein radii for high-dimensional disturbances scale poorly (O($u_k \in \mathbb{R}^m$4)), whereas MMD sets achieve dimension-independent O($u_k \in \mathbb{R}^m$5) rates (Nemmour et al., 2022).

Research Directions

• Extensions to higher-order convexification (e.g., DDP), φ-divergence ambiguity, or infinite-horizon problems.
• Kernel approximation techniques to lower computational cost in MMD-DRCCP.
• Integration with online learning for adaptive radius shrinking as more data become available.
• Hybridization of adaptive robust and distributionally-robust schemes for systems with mixed epistemic and aleatoric uncertainty (Gahlawat et al., 4 Sep 2025).

7. Summary Table of Key Methodological Features

Approach	Ambiguity Set	Tractability/Algorithm	Reference
Tube-based DR-MPC (linearized)	Wasserstein, moment	Convex QP/SOCP; feedback tightening	(Zhong et al., 2022, Zhong et al., 2023)
MMD-DRCCP	MMD kernel ball	RKHS/convex program; kernel methods	(Nemmour et al., 2022, Nemmour et al., 2021)
Moment-based cone tightening	Mean/covariance	Linear tightening; Calafiore–El Ghaoui	(Safaoui et al., 2021, Renganathan et al., 2022)
Adaptive covariance steering	Wasserstein (cert.)	L₁-adaptive + DR-CVaR/MILP	(Gahlawat et al., 4 Sep 2025)

In conclusion, distributionally-robust nonlinear steering combines data-driven and structural robustness, offering tractable and statistically rigorous safety envelopes for uncertain nonlinear systems with broad applicability to process control, robotics, and autonomous systems (Zhong et al., 2022, Zhong et al., 2023, Nemmour et al., 2022, Nemmour et al., 2021, Gahlawat et al., 4 Sep 2025).