To control a physical system—whether it is a quadcopter, a chemical reactor, or an unstable robotic pendulum—classical control theory requires us to write down its governing equations. Traditional frameworks like Model Predictive Control (MPC) or Linear Quadratic Regulators (LQR) rely entirely on these first-principles mathematical models.
However, real-world systems are rarely perfectly clean. Non-linearities, friction, unmodeled fluid dynamics, and environmental noise make deriving an exact analytical model incredibly difficult, if not impossible. When the physical model is wrong, classical controllers can fail catastrophically.
To bypass this bottleneck, a powerful research direction has emerged: data-driven control. Instead of forcing engineers to derive explicit physical equations, these techniques use machine learning, operator theory, and raw interaction data to synthesize controllers directly. Crucially, they do this while preserving the formal stability and safety guarantees that make classical control trustworthy.
Learning to Linearize
One of the most elegant tools in classical nonlinear control is feedback linearization. The core intuition is simple: if you know the exact nonlinearities of a system, you can design a control law that explicitly cancels them out, making the closed-loop system behave like a simple, predictable linear system.
When the dynamics are unknown, we can use data-driven models to learn this cancellation mechanism.
[Raw Nonlinear System] ---> (Learned Transformation) ---> [Virtual Linear System] ---> [Standard Linear Controller]
Raktim Gautam Goswami et al. demonstrate how to achieve this by pairing standard feedforward neural networks with Invertible Neural Networks (INNs). The INN learns the complex state and input transformations required to map a "black-box" nonlinear system into a controllable linear space, allowing engineers to apply standard linear control tools to highly complex systems. Similarly, Tyler Westenbroek et al. use model-free reinforcement learning to directly learn the linearizing control law, bridging the gap between data-driven optimization and exact trajectory tracking.
However, learning a static model offline can be risky if the system encounters novel environments. To address this, Sandra Hirche et al. combine feedback linearization with Gaussian Processes (GPs) that learn online. To prevent the computational overload of updating the GP at every time step, the authors introduce an event-triggered rule: the controller only requests a data update when model uncertainty crosses a specific threshold, guaranteeing real-time efficiency alongside formal stability.
Lifting Dynamics via Operators and Kernels
Another way to handle a nonlinear system without knowing its exact physics is to change our perspective. Instead of tracking how individual states evolve in a complex, low-dimensional nonlinear space, we can "lift" the system into an infinite-dimensional space of functions where the dynamics actually behave linearly. This is the core intuition behind operator-theoretic and kernel methods.
Nicolas Hoischen et al. apply this by learning the "infinitesimal generator" of a system's dynamics within a reproducing kernel Hilbert space (RKHS). This allows them to solve optimal control problems for complex, noisy, stochastic systems directly from data without ever needing a first-principles model. To avoid the tedious process of manually designing features for these spaces, Jiahang Liu et al. introduce Multi-Kernel Learning Control (MMLC), which automates feature engineering by combining multiple similarity kernels to learn optimal control policies for discrete-time systems.
Purely data-driven methods can sometimes suffer from low sample efficiency. If we happen to have some partial physical knowledge, we should use it. Adam Thorpe et al. introduce "physics-informed kernel embeddings," which seamlessly blend prior structural physics knowledge with data-driven kernel methods, drastically reducing the amount of data required to achieve precise control.
Direct Trajectory Optimization and Predictive Control
What if we don't try to learn a model or an operator at all? In many scenarios, we can use raw historical trajectories to implicitly represent what the system is capable of doing.
For linear systems, a foundational concept known as Willems' Fundamental Lemma proves that a single, sufficiently excited trajectory can fully characterize all possible behaviors of the system. Yifan Xie et al. exploit this to create a data-driven Economic Model Predictive Control (EMPC) scheme. By using raw input-output data instead of an explicit system model, their controller successfully guides unknown systems to their economically optimal operating steady states.
For highly unstable nonlinear systems, model-free reinforcement learning can find stabilizing policies directly through trial and error. Andrea Bonarini et al. demonstrate that Batch Reinforcement Learning (specifically Fitted Q-Iteration) can stabilize an unstable Mobile Wheeled Pendulum robot using as little as 100 seconds of random interaction data, ultimately outperforming classical LQR controllers and generalizing to entirely unseen movement profiles.
Baking Stability and Safety into Learning
The primary critique of pure machine learning in control loops is the lack of safety guarantees; a neural network can easily output an unpredictable value that destabilizes a physical machine. The modern frontier of data-driven control solves this by embedding classical control-theoretic constraints directly into the learning architectures.
[Neural Network Policy] ---> (Lyapunov / Safety Filter) ---> [Guaranteed Safe Action]
A prime example of this paradigm is Neural Identification for Control (NI4C), developed by Priyabrata Saha. Instead of training a standard neural network and hoping it is stable, the author embeds Lyapunov stability theory directly into the network's architecture. The system jointly learns the controller and a "stable hypothesis" of the closed-loop behavior, ensuring a vastly expanded and guaranteed safe operating region (Region of Attraction) compared to traditional reinforcement learning.
Taking safety a step further, Milad Kazemi focuses on formal synthesis, combining data-driven machine learning with formal logic. This framework allows for the automatic design of controllers that are mathematically guaranteed to satisfy strict safety specifications, even when operating in highly uncertain or completely unmodeled environments.
Key Foundational Work
Batch Reinforcement Learning for Controlling a Mobile Wheeled Pendulum Robot (2008) — contributed 1 concept elements to this direction.
Incremental approximate dynamic programming for nonlinear flight control design (2015) — contributed 1 concept elements to this direction.