Modern control engineering faces a fundamental tension. Empirical, data-driven methods like deep reinforcement learning can master highly complex, nonlinear systems without requiring an explicit mathematical model (Spielberg et al.). However, these black-box algorithms often lack the rigorous safety nets and closed-loop stability proofs that classical control theory provides (Liu et al.). In safety-critical applications—such as aviation, robotics, or chemical processing—an uncertified controller can cause catastrophic physical failure.
To resolve this, a major research direction focuses on embedding systems-theoretic guarantees directly into data-driven frameworks (Allgöwer et al.). Instead of treating data-driven models as unconstrained black boxes, researchers use foundational control principles to bound uncertainty, guarantee closed-loop stability, and ensure constraint satisfaction.
Predicting Trajectories Directly from Data
Traditional indirect control requires a two-step process: collect data, fit a parametric model, and then design a controller for that model. However, modeling errors can compound, destroying the theoretical safety guarantees of the resulting controller.
Direct data-driven Model Predictive Control (MPC) bypasses model identification entirely. The core intuition relies on behavioral systems theory: for a linear time-invariant system, any valid input-output trajectory can be represented as a linear combination of previously collected historical data trajectories, provided the historical data is sufficiently exploratory.
Berberich et al. leveraged this insight to formulate a purely data-driven MPC scheme that provides rigorous closed-loop stability and robustness guarantees. By embedding raw historical data directly into the MPC optimization problem as a non-parametric predictor, the controller implicitly respects the underlying physical laws of the system. In subsequent work, the authors validated this approach experimentally, proving that direct data-driven control can maintain strict safety constraints without an explicit mathematical model (Berberich et al.). To extend this concept to tracking tasks in nonlinear regimes, Berberich et al. developed a tracking MPC framework that allows unknown nonlinear systems to safely follow reference trajectories using only recent local history.
To understand why direct data-driven control works so robustly, Coulson et al. established a mathematical bridge between direct and indirect methods. They proved that direct data-driven control formulations (such as Data-enabled Predictive Control, or DeePC) are actually convex relaxations of the classical indirect approach. This explains their inherent resilience to noise and provides a systematic way to design regularizers that enforce stability.
Structuring the Learning Space with Stability Certificates
When dealing with highly nonlinear or high-dimensional systems, linear behavioral theory is no longer sufficient. If we want to use flexible function approximators like neural networks, how do we prevent them from generating destabilizing control actions?
One approach is to structure the controller's parameterization or the exploration process using classical Lyapunov stability theory. A Lyapunov function acts as a mathematical "energy" metric; if a controller guarantees that this energy strictly decreases over time, the system is guaranteed to converge safely to its equilibrium.
Instead of guessing a Lyapunov function manually, Chang et al. proposed training a neural network called an Almost Lyapunov Critic alongside a model-free reinforcement learning policy. The critic self-learns a stability certificate, allowing the system to estimate its Region of Attraction and guarantee safety without prior knowledge of the dynamics. Similarly, Fan et al. introduced Lyapunov-Guided Exploration (LYGE), which uses Lyapunov boundaries to restrict data collection exclusively to states that are known to be safe, drastically reducing the sample complexity needed to stabilize high-dimensional unknown systems.
Alternatively, we can constrain the policy space itself so that every candidate controller is stable by design. Barbara et al. introduced the "Youla-REN" framework, which combines a nonlinear extension of the Youla parameterization with Recurrent Equilibrium Networks (RENs). Because the structural properties of RENs inherently satisfy contraction and Lipschitz bounds, any neural network trained within this framework is mathematically guaranteed to yield a stable closed-loop system. For linear systems, Holicki et al. proposed a related parameterization strategy that restricts candidate controllers to a safe, stabilizing subspace, allowing safe experimental exploration and tuning directly on physical hardware.
For systems with extreme dimensionality, such as soft continuum robots, structuring the space requires reducing the system's degrees of freedom. Alora et al. utilized Spectral Submanifold (SSM) theory to reduce thousands of degrees of freedom to a handful of dominant dimensions, enabling safe, precise, and real-time model-based tracking.
Quantifying Finite-Data Uncertainty and Robustness
In practice, data is never perfect; it is finite, noisy, and subject to distribution shifts. To provide deterministic guarantees, a controller must know exactly how much to trust its data.
Boczar et al. addressed this by establishing finite-data performance bounds for output-feedback control. Their framework quantifies the exact relationship between the volume of collected data and the resulting robust control performance, giving engineers a clear metric for data sufficiency. To handle noise and nonlinearities simultaneously, Huang et al. introduced a kernelized data-driven predictive control algorithm (RoKDeePC). By mapping input-output data into a high-dimensional kernel space and wrapping the predictor in a min-max optimization framework, they achieve deterministic performance guarantees despite measurement noise.
When a system must adapt online, controllers must balance exploration with caution. Wabersich developed Cautious Bayesian MPC, proving that by systematically tightening constraints based on epistemic uncertainty during the learning process, one can strictly bound the number of unsafe learning episodes. To reduce the conservatism of these bounds over time, Aboudonia et al. combined distributed set-membership identification with an adaptive tube MPC approach, allowing the controller to safely learn unknown subsystem couplings online. Finally, Liu et al. closed the loop between statistical learning and physical execution by using a minimax optimization lens to map regression generalization errors directly to physical tracking performance bounds.
Key Foundational Work
Finite-Data Performance Guarantees for the Output-Feedback Control of an Unknown System (2018) — contributed 1 concept elements to this direction.
Stability of Controllers for Gaussian Process Dynamics (2017) — contributed 1 concept elements to this direction.