In traditional control engineering, the path from a physical system to an autonomous controller is a two-step process: first, engineers derive a mathematical model (system identification), and then they design a controller based on that model. However, for complex systems like soft robots, multi-agent legged loccomotion, or modern power grids, building an accurate model is often the most difficult and error-prone part of the pipeline.
A growing body of research suggests we can skip the modeling step entirely. By leveraging Behavioral Systems Theory, researchers have developed Data-EnablEd Predictive Control (DeePC), a framework that uses raw historical trajectories to predict and control future behavior directly.
The Intuition: Data as a Non-Parametric Model
The core insight of this field rests on Willems’ Fundamental Lemma. As explained by Ivan Markovsky and Linbin Huang in Data-Driven Control Based on the Behavioral Approach, for a linear system, any possible trajectory can be represented as a linear combination of a sufficiently "rich" set of past trajectories.
Think of it as a library of behaviors. If you have recorded enough diverse examples of how a robot moves, any new valid movement must be a "patchwork" of those recorded examples. Instead of solving for physical parameters like mass or friction, DeePC solves for the coefficients that combine these past experiences to reach a desired future state.
Bridging the Gap to Nonlinearity
While the Fundamental Lemma was originally for linear systems, researchers like Mohammad Alsalti, Julian Berberich, and Matthias A. Müller have extended this logic to nonlinear systems. In Data-based Control of Feedback Linearizable Systems, they demonstrate that systems which can be mathematically transformed into linear ones (feedback linearizable) can also be controlled using data-driven basis functions.
For even more complex "squishy" systems, such as the 3D soft robotic arm studied by Moeen Ul Islam and Xiaobo Tan in Direct Data-Driven Predictive Control for a Three-dimensional Cable-Driven Soft Robotic Arm, DeePC bypasses the "nightmare" of modeling soft-body dynamics, achieving precise tracking by simply observing how the cables affect the arm's position over time.
Robustness through Regularization
In the real world, data is never perfect—it is plagued by sensor noise and stochastic disturbances. If a controller tries to perfectly match a new trajectory to noisy historical data, it will "overfit," leading to erratic and unstable control actions.
The solution, appearing across nearly all recent literature, is the integration of regularization and slack variables.
- Regularization: By adding penalty terms to the optimization objective, the controller is discouraged from using overly complex combinations of past data. This acts as a "noise filter," as shown in Distributed Data-Driven Predictive Control for Multi-Agent Collaborative Legged Locomotion.
- Slack Variables: These act as buffers. If the current sensor reading doesn't perfectly align with the historical "behavioral manifold," slack variables allow the optimization to find a "close enough" solution rather than failing. Mohammad Alsalti et al. use these to prove practical exponential stability in Practical exponential stability of a robust data-driven nonlinear predictive control scheme, ensuring the system stays within a safe boundary even with imperfect data.
Real-World Impact: From Power Grids to Batteries
The power of this approach is its adaptability to systems that change over time. In Degradation-aware data-enabled predictive control of energy hubs, Marta Zagorowska and John Lygeros show that DeePC can manage building energy hubs and double battery lifespans by learning from real-time degradation data, rather than relying on a static (and eventually inaccurate) battery model.
Similarly, in power electronics, Ruohan Leng and Linbin Huang developed DeePConverter, which allows grid-connected converters to "perceive" and adapt to unknown grid dynamics. By replacing traditional PID loops with data-driven predictive control, these systems become significantly more robust to the fluctuations of modern renewable energy grids.
Summary
DeePC represents a shift from modeling physics to curating behavior. By treating a library of past trajectories as the model itself and using regularization to navigate the noise, researchers are creating controllers that are not only easier to design but more resilient to the complexities of the physical world. As John Lygeros and Florian Dorfler demonstrate with RoKDeePC, this "model-free" future provides deterministic performance guarantees that were once thought possible only with perfect mathematical equations.