To control a physical system—whether it is a robotic arm, a chemical reactor, or an industrial boiler—traditional control theory requires a mathematical map. This map, typically a set of differential equations, dictates how a given input (like a voltage or a valve opening) alters the system's state over time.
However, real-world systems are notoriously messy. They exhibit severe nonlinearities, unmodeled friction, time delays, and complex cross-coupling between variables. As Jiang Xinsong et al. observed, a stark discrepancy often exists between simplified model assumptions and real-world environmental variables, leading to poor control performance. Deriving these models from first principles of physics is frequently too labor-intensive or mathematically intractable. Frank Allgöwer identified this lack of systematic modeling methodologies as a major bottleneck preventing advanced control theories from being adopted in industrial settings.
If we cannot easily build an accurate physical model, how do we safely and effectively control complex nonlinear systems? A powerful research direction has emerged to bypass this modeling bottleneck entirely, shifting from model-based control to data-driven and model-free control paradigms.
The Danger of Explicit System Identification
The intuitive fallback when physics fails is statistical system identification: using data to fit a complex nonlinear model, and then designing a controller for that fitted model. But this two-step approach has a major flaw. As Longquan Ma et al. point out, directly identifying unknown nonlinear dynamics is highly impractical and prone to compounding errors; a small mismatch in the model's equations can cause the controller to become unstable when deployed on the actual hardware.
Furthermore, traditional system identification rarely quantifies its own errors. Without rigorous error bounds, engineers cannot use robust control tools that guarantee safety. To address this, Tim Martin et al. developed a framework to calculate a system's "strength of nonlinearity" and guaranteed error bounds directly from noisy data, avoiding the need to construct a full mathematical model first.
Instead of trying to force a messy reality into an explicit, rigid set of nonlinear equations, modern approaches focus on three main strategies: local adaptation, linear lifting, and direct data-to-control optimization.
Strategy 1: Model-Free Adaptive Control
If you are walking up a mountain in dense fog, you do not need a complete topographic map of the peak to move upward. You only need to feel the slope right beneath your feet and adjust your step accordingly.
This is the intuition behind Model-Free Adaptive Control (MFAC). Instead of modeling the global dynamics of a system, MFAC continuously estimates a local, dynamic relationship between changes in the control input and changes in the system output—a concept known as a pseudo-gradient.
[Input/Output Data Stream] ---> [Dynamic Parameter Tuning (LSTM/FNN)]
|
v
[System State] ------------> [Local Pseudo-Gradient Estimation] ---> [Control Action]
Because these pseudo-gradients change over time, the controller must adapt rapidly. To make this process robust against noise and sudden system shifts, Ye Yang et al. integrated Long Short-Term Memory (LSTM) neural networks to handle parameter self-tuning on the fly. By expanding this framework to full-form dynamic linearization, Ye Yang et al. combined LSTMs with Fuzzy Neural Networks to estimate these gradients more accurately, allowing model-free controllers to track trajectories in highly unpredictable environments without ever knowing the underlying physics.
Strategy 2: Lifting Nonlinearity into Linearity
Linear systems are incredibly well-behaved. If a system is linear, we can use classic, mathematically elegant tools like the Linear Quadratic Regulator (LQR) to find the mathematically optimal control strategy. Unfortunately, most interesting real-world systems are nonlinear.
Koopman operator theory provides a remarkable mathematical loophole: any nonlinear system can be represented as a linear system, provided you map its states into an infinite-dimensional space of "observables."
Imagine trying to track the complex, nonlinear motion of a shadow cast on a rugged, bumpy wall. If you look only at the 2D coordinates of the shadow, the math is horrific. But if you "lift" your perspective to the 3D object casting the shadow, its motion might be a simple, linear rotation.
Low-Dimensional Nonlinear Space High-Dimensional Linear Space
[ x_t = f(x_{t-1}) ] [ z_t = A * z_{t-1} ]
| ^
| (Lift via Observables) | (Apply Linear Control: LQR)
v |
[ State: x ] =======================> [ Lifted State: z ]
To operationalize this concept, Carl Folkestad et al. introduced Koopman Eigenfunction Extended Dynamic Mode Decomposition (KEEDMD). This method uses data to learn the specific dictionary of functions (eigenfunctions) needed to lift a robotic or physical system into a linear state-space model.
Once the system is lifted, control becomes straightforward. Max Beier et al. demonstrated that by leveraging these "Koopmanizing Flows," complex nonlinear optimal control problems can be seamlessly transformed into standard linear LQR problems. To ensure that the inevitable errors introduced by truncating an infinite-dimensional space do not cause instability, Robin Strässer et al. combined Koopman surrogates with Sum-of-Squares (SOS) optimization, providing rigorous stability guarantees directly from data.
Strategy 3: Direct Data-Enabled Predictive Control
Model Predictive Control (MPC) is an industry-standard technique that works by solving an optimization problem at every time step: it uses a model to predict future states over a time horizon, chooses the best sequence of inputs, executes the first step, and repeats.
Can we perform MPC without the model? A foundational principle in data-driven control (often leveraging Fundamental Lemmas like Willems' Fundamental Lemma) states that a single, sufficiently exciting historical trajectory of a linear system contains all the information necessary to predict any other future trajectory.
Julian Berberich et al. extended this philosophy to nonlinear tracking, proving that a system can effectively predict its future and control itself on the fly using only its recent history, bypassing explicit differential equations entirely.
[Historical Trajectories Matrix]
+ ---> [Min-Max Optimization] ---> [Optimal Control Input]
[Current Recent History]
When data is noisy, relying directly on raw historical trajectories can be dangerous. To fix this, Linbin Huang et al. introduced Robust and Kernelized Data-Enabled Predictive Control (RoKDeePC). By using kernel methods, they implicitly map input-output data into high-dimensional feature spaces to capture nonlinearities, combining this with a min-max optimization framework to guarantee deterministic performance even when the data is corrupted by noise.
Bridging the Gap
By moving away from rigid first-principles models, modern control engineering is finding ways to handle systems that were previously unmanageable. Whether through reinforcement learning frameworks that learn linearizing controllers Tyler Westenbroek et al., or neural networks that diagnose faults without physical models Haipeng Pan, the trend is clear. By treating data not as a tool to build static models, but as a dynamic representation of the system itself, these techniques provide the mathematical rigor of classical control theory without the crippling bottleneck of physical system identification.
Key Foundational Work
A hierarchical approach to designing approximate reasoning-based controllers for dynamic physical systems (1990) — contributed 1 concept elements to this direction.
AI Research in China: A Review (1983) — contributed 1 concept elements to this direction.
AI Research in China: A Review (1983) — contributed 1 concept elements to this direction.