The Motor That Taught Itself to Stay Stable — and Beat the Kalman Filter by 35%

Somewhere inside an electric vehicle, a robotic arm, or an industrial pump, a motor is spinning. It is being controlled by software making thousands of decisions per second — adjusting current, anticipating load, correcting for tiny delays in sensing and actuation. Most of the time, this works well enough. But "well enough" is doing a lot of work in that sentence. Real motors operate under time-varying delays, mechanical disturbances, and model uncertainty. Standard control methods — even sophisticated ones — often handle these challenges through workarounds rather than guarantees.

A new paper by Syed Pouladi (Pouladi, 2026) proposes something more principled: a control framework for electric motors that is stable by mathematical construction, learns its model from real trajectory data, and outperforms state-of-the-art baselines by a substantial margin. On a 1.5 kW Permanent Magnet Synchronous Motor (PMSM) — the kind of motor found in electric vehicles, servo drives, and precision manufacturing — the framework reduced velocity estimation error by 35% compared to an Extended Kalman Filter, and improved speed-tracking accuracy by 67% relative to standard Field-Oriented Control (FOC), the industry workhorse. Those are not marginal gains. They suggest that a fundamentally different approach to motor control is possible.

The Science

The paper operates at the intersection of three fields that rarely speak to each other this fluently: nonlinear dynamical systems theory, data-driven machine learning, and optimal control. The unifying thread is a class of mathematical models called generalized Persidskii systems.

A Persidskii system is a way of writing down the equations of a physical system — a motor, a circuit, a neuron — that separates the linear parts of the dynamics from the nonlinear parts. The nonlinear parts are constrained to be sector-bounded, meaning their outputs always fall within a cone-shaped region around zero. This is a mild but useful restriction: it rules out catastrophically explosive nonlinearities while still capturing a wide range of real behavior. Electric motors, it turns out, fit naturally into this framework. So do many neural network models, which is why Persidskii systems have attracted renewed theoretical interest.

The key stability concept the paper develops is input-to-state stability (ISS) — a property that says, roughly, if the disturbances hitting a system are bounded, the system's states will also remain bounded. ISS is a strong guarantee: it tells you the system won't blow up even when things go wrong. To derive ISS conditions for Persidskii systems with time-varying delays, Pouladi uses Lyapunov–Krasovskii functionals — generalized energy functions that account for the history of the system's states, not just their current values. This is essential for delay systems, where what happened a few milliseconds ago can determine whether the system stays stable today.

These conditions are then expressed as Linear Matrix Inequalities (LMIs) — a class of convex optimization problems that can be solved efficiently and reliably with standard software. LMIs are the lingua franca of modern robust control: if you can write your stability condition as an LMI, you can check it, enforce it, and design around it.

The framework also includes a structured robust observer — a software system that estimates the motor's internal states (like rotor flux and velocity) from partial measurements (typically only current and voltage are directly available). The observer's convergence is guaranteed via an $H_{\infty }$ synchronization criterion, which ensures that errors between estimated and true states shrink even in the presence of noise and disturbances. The $H_{\infty }$ framework, named for the mathematical space in which it operates, is the gold standard for robust estimation: it minimizes the worst-case amplification of disturbances through the observer.

The motor's mathematical model is not assumed to be perfectly known in advance. Instead, it is identified from data — trajectory measurements collected from the actual motor — using a Koopman lifting procedure. The Koopman operator is a concept from dynamical systems theory that allows a nonlinear system to be represented, approximately, as a linear system in a higher-dimensional space. Think of it as finding a coordinate system in which the complicated, curved dynamics of a motor look straight. The lifting maps the original states into a richer feature space; the linear dynamics in that space are then fitted to data.

The critical innovation here is how the fitting is done. Rather than learning the best model and then checking whether it is stable, Pouladi embeds the ISS-LMI constraint directly into the regression problem as a convex side condition. Stability is not an afterthought — it is baked into the objective. The result is a learned model that is guaranteed, by construction, to satisfy the same stability properties that the theoretical analysis demands.

This identified model then feeds into an ICODE-MPPI controller — a mouthful of an acronym for Input-dependent Control-oriented Dynamical Estimation combined with Model Predictive Path Integral control. MPPI is a sampling-based optimal control method that rolls out thousands of possible future trajectories simultaneously (on a GPU, typically) and selects the control input that minimizes expected cost. It is computationally intensive but powerful: it handles nonlinearities, constraints, and complex cost functions gracefully, without requiring linearization. The "ICODE" component adapts the prediction model based on the current operating point, making the controller responsive to changing conditions.

What They Found

The experimental results are the paper's most striking contribution. The complete framework was validated on a 1.5 kW PMSM drive equipped with a programmable load brake — a device that can impose known, controlled mechanical loads on the motor shaft, simulating real operating conditions including sudden disturbances.

Two comparisons anchor the results. First, the velocity estimation performance of the proposed Persidskii observer against an Extended Kalman Filter (EKF) — arguably the most widely used state estimator in motor drives, and a strong baseline. The proposed observer achieved a 35% reduction in root-mean-square error (RMSE) for velocity estimation. RMSE is the standard measure of estimation accuracy: it captures the average magnitude of the error between estimated and true values, penalizing large errors more heavily. A 35% reduction is substantial in a domain where incremental improvements are the norm.

Velocity Estimation RMSE: Proposed Observer vs. Extended Kalman Filter

Label	Value
Extended Kalman Filter (baseline)	100 %
Persidskii Observer (proposed)	65 %

35% RMSE reduction. Source: Pouladi (2026), arXiv:2604.27509. EKF set to 100% for relative comparison.

Second, speed-tracking accuracy under the ICODE-MPPI controller was compared against Field-Oriented Control (FOC), the dominant method in industrial PMSM drives. FOC works by decomposing motor currents into torque-producing and flux-producing components and controlling them independently — a clever transformation that makes a nonlinear system behave, approximately, like a linear one. It is effective, well-understood, and deeply embedded in industrial practice. Against this benchmark, the proposed framework achieved a 67% improvement in speed-tracking accuracy.

Speed-Tracking Accuracy: ICODE-MPPI vs. Field-Oriented Control

Label	Value
Field-Oriented Control (baseline)	100 %
ICODE-MPPI (proposed)	33 %

67% improvement in speed-tracking accuracy. Source: Pouladi (2026), arXiv:2604.27509. FOC set to 100% for relative comparison.

These two numbers together tell an important story. The 35% improvement in estimation is not merely academic — better state estimates directly enable better control. And the 67% improvement in tracking confirms that the benefits propagate all the way through the system, from the observer to the controller to the actual motor shaft. The theoretical ISS bounds established in the paper, the authors note, are corroborated by the experimental results: the system behaved, in practice, exactly as the mathematics predicted it should.

Why This Changes Things

To understand why this matters, it helps to appreciate the gap that usually exists between control theory and control practice. Theorists develop stability conditions, prove convergence theorems, and derive guarantees — but these results often rely on assumptions (perfect model knowledge, no delays, bounded disturbances of a specific form) that are violated in any real system. Practitioners, meanwhile, tune PID controllers and Kalman Filters empirically, building in safety margins through experience rather than proof.

The framework in this paper is unusual because it genuinely closes part of that gap. The stability guarantees are not just theoretical ornaments — they are directly enforced during the learning process and reflected in the experimental outcomes. This is the deeper significance of embedding the LMI constraint into the Koopman regression: it means that the learned model inherits the same formal properties that the theory demands. You get the expressiveness of a data-driven model and the reliability of a theoretically certified one.

This matters especially for time delays. Delays are ubiquitous in real systems: sensor sampling introduces them, communication networks introduce them, actuator dynamics introduce them. A controller that is stable without delays can become unstable when delays are added — a phenomenon responsible for many real-world control failures. The paper's delay-dependent ISS conditions account for this explicitly, providing certificates that remain valid even as delay lengths vary over time.

The Persidskii framework also has breadth beyond motor control. As Pouladi notes, recurrent neural networks and many other machine learning architectures can be written in Persidskii form. The sector-bounded nonlinearity condition is satisfied by standard activation functions like $\tanh$ and sigmoid. This means the same stability analysis, observer design, and data-driven identification machinery could, in principle, be applied to neural network dynamical models — a connection with implications for safe AI systems that must operate in physical environments.

The MPPI controller brings its own advantages. Because it is sampling-based rather than gradient-based, it handles non-smooth cost functions and hard constraints naturally. It does not require the dynamics to be differentiable with respect to the control input at every point. And because it runs on parallel hardware, it can evaluate thousands of candidate trajectories in the time it takes a classical solver to evaluate one. The combination of a theoretically certified model with a powerful nonlinear optimizer is, arguably, what produces the large tracking improvement the experiments demonstrate.

For electric motor drives specifically, the stakes are high. PMSMs are in electric vehicles, wind turbines, industrial robots, aerospace actuators, and home appliances. Global motor-driven systems consume roughly 45% of the world's electricity. Even modest improvements in control efficiency — by reducing the energy wasted on suboptimal tracking, or by enabling faster, more precise response — aggregate into substantial real-world impact at scale.

What's Next

The paper is, by its own framing, a proof of concept at the scale of a single 1.5 kW drive. The natural next questions are about generalization: Does the framework scale to larger, more complex systems? How does the Koopman lifting procedure perform when the dynamics are higher-dimensional or more strongly nonlinear? How many trajectory samples are needed to produce a reliable model, and how sensitive are the results to data quality?

There are also open questions about the MPPI controller. MPPI's computational demands grow with the number of sample trajectories and the length of the prediction horizon. On a GPU-equipped embedded system, this is increasingly feasible — but the paper does not detail the computational budget required for real-time operation, which is a practical concern for deployment in resource-constrained industrial hardware.

The connection to neural network systems is tantalizing but underdeveloped in this work. A follow-on study that explicitly applies the Persidskii stability conditions to a recurrent neural network used as a plant model — or as a component of the controller itself — could open a genuinely new line of research in certifiable machine learning for dynamical systems.

The observer design also warrants further investigation. The $H_{\infty }$ criterion guarantees robustness against worst-case disturbances, but real disturbance distributions may be far from worst-case. A Bayesian or probabilistic extension of the observer — one that takes advantage of known noise statistics when they are available, while falling back to $H_{\infty }$ guarantees when they are not — could extract additional estimation performance.

Finally, the paper validates the framework under controlled laboratory conditions with a programmable load brake. Real industrial environments introduce additional challenges: temperature variation, bearing wear, magnetic saturation, electromagnetic interference. Validation under these conditions, or explicit modeling of their effects within the Persidskii framework, would be a necessary step toward deployment.

What this paper establishes, even within those caveats, is meaningful: that it is possible to build a control system for a real physical plant that is simultaneously learned from data, certified to be stable by construction, and substantially better in practice than the methods it replaces. For a field that has long lived with the tension between theoretical elegance and practical utility, that is not a small thing. The mathematics and the motor, for once, agreed.