⚡ electrical engineering

DUST: A Framework for Data-Driven Density Steering

This paper introduces DUST, a novel framework that leverages noisy data from an unknown LTI system to simultaneously estimate noise characteristics and design optimal affine-feedback controllers for steering the state's mean and covariance to a target distribution using robust control and optimization techniques.

Original authors: Joshua Pilipovsky, Panagiotis Tsiotras

Published 2026-02-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Joshua Pilipovsky, Panagiotis Tsiotras

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guide a very clumsy, invisible robot from Point A to Point B. You don't have a manual for the robot (you don't know its exact physics), and the robot is constantly being bumped by invisible, random gusts of wind (noise). Your goal isn't just to get the robot to the exact center of Point B; you want to ensure that the entire cloud of possible robot positions ends up in a specific, neat shape at Point B.

This paper, titled DUST, presents a new "recipe" for doing exactly that using only a history of past attempts (data), without ever needing to know the robot's internal blueprint.

Here is a breakdown of how the authors' framework works, using everyday analogies:

1. The Problem: The "Blindfolded" Pilot

Traditionally, to control a system, you first build a perfect model (a blueprint) and then design a controller. But in the real world, building a perfect blueprint is hard, and the data you collect is always "noisy" (like trying to hear a whisper in a storm).

The Old Way: Try to guess the blueprint, then fly. If your guess is wrong, you crash.
The DUST Way: Skip the blueprint. Look at the raw data of past flights, figure out how the wind actually behaved, and design a controller that is robust enough to handle the uncertainty.

2. The Core Idea: Steering a Cloud, Not a Dot

Most control problems try to steer a single point (the average position). This paper treats the system as a cloud of possibilities (a probability distribution).

The Analogy: Imagine you are throwing a handful of marbles. You want all the marbles to land in a specific small circle at the end.
- Mean Steering: You want the center of the marble pile to land in the middle of the circle.
- Covariance Steering: You want the spread of the marbles to be tight enough that none spill out of the circle.
- DUST does both at the same time, ensuring the whole cloud lands exactly where you want it.

3. The "Noise Detective" (Noise Estimation)

Since the data is noisy, the authors realized they can't just trust the raw numbers. They need to act like detectives to figure out what the "wind" (noise) actually did during the past experiments.

The Method: They use a statistical technique called Maximum Likelihood Estimation (MLE). Think of this as a detective looking at the crime scene (the data) and saying, "Given what I see, this is the most likely sequence of wind gusts that happened."
The Twist: They don't just guess the wind; they calculate a "confidence bubble" around their guess. They know, "I'm 95% sure the wind was this strong, give or take a little bit."

4. The Safety Net (Robust Control)

Once they have their "most likely" wind history and a "confidence bubble" for how wrong they might be, they design a controller that works for every scenario inside that bubble.

The Analogy: Imagine you are packing for a trip. Instead of packing for "average weather," you pack for "average weather plus a potential storm."
The Result: Even if their guess about the wind was slightly off, the controller is strong enough to steer the cloud of marbles to the target anyway. This is called Robust Control.

5. Two Different Approaches

The paper offers two main ways to solve this puzzle:

The "Indirect" Way (The Detective): First, estimate the wind and the robot's behavior separately. Then, build a controller that is extra careful to handle any mistakes in that estimation. This is very safe but requires complex math to prove it works.
The "Parametric" Way (The Gambler): Instead of guessing the specific wind gusts, treat the wind as a known type of random variable (like a known dice roll). You design a controller that is statistically likely to work for the entire distribution of possible winds. This is simpler but relies on the assumption that the wind follows a specific pattern (like a bell curve).

6. The Results: Does it Work?

The authors tested their "DUST" framework on a simulated robot.

Comparison: They compared it to the "Standard Way" (guessing the blueprint and ignoring the noise) and found that the Standard Way often missed the target or spread the marbles too wide.
Success: The DUST framework successfully steered the cloud of marbles to the target with high precision, even when the data was very noisy. They showed that by using their "confidence bubbles," they could guarantee the robot would land safely, whereas other methods failed.

Summary

DUST is a framework that admits, "We don't know the exact rules of the game, and our data is messy." Instead of panicking, it uses statistics to figure out the most likely messiness, creates a safety margin around that guess, and designs a controller that is guaranteed to win the game even if the messiness is slightly different than expected. It turns a chaotic, noisy problem into a solvable, safe one.

Technical Summary: DUST – A Framework for Data-Driven Density Steering

Problem Statement
The paper addresses the problem of data-driven stochastic optimal control for unknown Linear Time-Invariant (LTI) systems subject to Gaussian process noise. The objective is to steer the probability density of the system state from a known initial Gaussian distribution to a prescribed terminal Gaussian distribution. This is formulated as a Covariance Steering (CS) problem, where the goal is to control both the mean and the covariance of the state trajectory.

Unlike traditional model-based approaches that first identify a system model and then design a controller, or existing direct data-driven methods that often assume deterministic dynamics or noise-free data, this work tackles the challenge of designing optimal controllers using only raw, noisy input-state data collected offline. The system matrices ( $A, B, D$ ) are unknown, and the data is corrupted by stochastic disturbances. The core challenge lies in accounting for the uncertainty introduced by noisy data collection while ensuring that the terminal distribution constraints are satisfied with high probability.

Methodology
The proposed framework, termed DUST (Data-driven Uncertainty quantiﬁcation and density STeering), decomposes the problem into two sub-problems: Data-Driven Mean Steering (DD-MS) and Data-Driven Covariance Steering (DD-CS). The authors leverage behavioral systems theory (specifically Willems' Fundamental Lemma) and statistical learning to parameterize the control laws directly from data.

Data-Driven Parameterization:
- Indirect Approach (DD-MS): The mean dynamics are parameterized using an indirect design. The system matrices are estimated via a least-squares approach (Certainty Equivalence) using the collected data, resulting in an uncertain quadratic program where the uncertainty stems from the unknown noise realization.
- Direct Approach (DD-CS): The feedback gains for covariance control are parameterized directly using the input-state data and subspace identification. This leads to a Semi-Definite Program (SDP) formulation.
Noise Estimation and Uncertainty Quantification:
- Since the noise realization ( $\Xi_{0,T}$ ) and its covariance ( $\Sigma_\xi$ ) are unknown, the framework employs Maximum Likelihood Estimation (MLE) to estimate the most probable noise realization from the dataset.
- Crucially, the paper derives high-confidence ambiguity sets (uncertainty bounds) for the estimation errors. Unlike previous works that used overly conservative bounds, this paper utilizes the statistical properties of the MLE estimator and the rank-deficiency of the error covariance matrix to construct "tight" ellipsoidal uncertainty sets. These sets guarantee that the true noise realization lies within the bounds with a prescribed probability ( $1-\delta$ ).
Robust Control Synthesis:
- Robust DD-MS: The mean steering problem is reformulated as a robust control problem with unstructured model uncertainty. Using System Level Synthesis (SLS), the authors transform the semi-infinite robust constraints into a tractable convex optimization problem (SDP) that guarantees the terminal mean constraints are satisfied for all uncertainties within the derived bounds.
- Robust DD-CS: The covariance steering problem is reformulated to robustly satisfy Linear Matrix Inequality (LMI) constraints over the planning horizon, accounting for the uncertainty in the noise realization.
- Parametric Uncertainty (PU-DD-DS): As an alternative, the paper presents a formulation treating the noise as a stochastic variable with known distribution (multiplicative uncertainty) rather than estimating specific realizations. This approach avoids explicit noise estimation but relies on a Certainty Equivalence assumption for the mean dynamics.

Key Contributions

Unified Framework: Introduction of a unified data-driven framework for steering the full state distribution of an unknown stochastic LTI system, combining behavioral systems theory with model-based covariance steering concepts.
Noise Estimation Tools: Development of data-driven noise estimation and uncertainty quantification tools that provide finite-sample, high-confidence ambiguity sets for estimation errors, improving upon previous conservative bounds.
Tractable Formulations: Provision of tractable (convex) formulations for both DD-MS and DD-CS problems. These formulations guarantee terminal distribution constraint satisfaction with high probability.
Parametric Uncertainty Approach: Presentation of a Parametric Uncertainty (PU-DD-DS) formulation that interprets the problem as controlling a system under multiplicative uncertainty, offering an alternative to the noise-estimation-based approach.
Performance Evaluation: Comprehensive numerical evaluation comparing the proposed data-driven designs against Certainty Equivalence (CE) baselines and model-based counterparts.

Results
Numerical experiments on a linear stochastic system demonstrate the efficacy of the DUST framework:

Noise Estimation: The proposed MLE-based uncertainty sets (specifically the "Tight" subspace decomposition) closely match the empirical quantiles of the true estimation error, whereas previous "Loose" bounds were significantly over-conservative and degraded with larger sampling horizons.
Robustness: The robust DD-MS design significantly outperforms the standard CE design in terms of controlling the dispersion of terminal mean trajectories, particularly when data is sparse or the model is uncertain.
Feasibility: The robust covariance design (R-DD-CS) is shown to be essential for feasibility; without accounting for estimation errors, the terminal covariance constraints are frequently violated. However, the paper notes a trade-off: increasing the robustness level ( $\rho$ ) or decreasing the terminal constraint size can lead to infeasibility if the theoretical lower bound on achievable covariance is approached.
Closed-Loop Performance: The combined robust mean and covariance design (R-DD-DS) achieves the closest terminal distribution to the desired target compared to other data-driven and model-based baselines in the presence of noisy data.

Significance and Claims
The paper claims that DUST provides a principled and rigorous approach to data-driven stochastic control, bridging the gap between behavioral system theory and robust control. Its primary significance lies in:

Handling Noisy Data: Moving beyond the assumption of deterministic or noise-free data, which is a major limitation of the Fundamental Lemma in real-world applications.
Tight Uncertainty Bounds: Offering a method to construct uncertainty sets that are significantly less conservative than prior art, thereby enabling more precise control without sacrificing safety guarantees.
End-to-End Solution: Providing a direct path from raw data to a robust controller that steers the state distribution, bypassing the need for explicit system identification that may fail for complex models.

The authors acknowledge limitations, noting that the current theory assumes linear time-invariant dynamics and Gaussian disturbances. They identify future work directions, including the incorporation of chance constraints, handling distributional uncertainty (e.g., Wasserstein ambiguity sets), and extending the framework to nonlinear systems or deriving a stochastic version of the Fundamental Lemma for mean and covariance trajectories.