Adaptive Control without Persistent Excitation

This direction is merged into another direction: persistent excitation requirement. Its last version stays readable and citable here.

Version 1 of 120 papers21 conceptsfrozen 2026-10-04

In classical control theory, if you want a robot to learn its own physical properties—like the friction in its joints or the weight of a tool it’s holding—you usually have to shake it. This requirement is known as Persistence of Excitation (PE). To identify unknown parameters, the system must be "excited" by signals that are sufficiently rich and varied. Without this constant wiggling, the math behind traditional adaptive controllers simply fails to converge on the truth.

However, in the real world, shaking is often dangerous. A stroke rehabilitation robot shouldn't jerk a patient's arm to learn its motor constants Borna Ghannadi and John McPhee note that such "persistently exciting" movements are at odds with smooth, therapeutic motion. Similarly, a power grid cannot be intentionally destabilized just to help a controller learn its dynamics Linbin Huang and Florian Dorfler explain that such disturbances compromise service quality.

A growing body of research is developing ways to bypass this PE requirement, allowing systems to learn and adapt while maintaining smooth, safe, and steady trajectories.

The Intuition of Concurrent Learning

The most prominent solution to the PE problem is Concurrent Learning (CL). To understand why it works, consider how a standard adaptive controller "forgets." Traditional algorithms update their estimates based only on the current instantaneous error. If the system stops moving in a way that reveals a specific parameter, the controller loses the "pressure" needed to maintain an accurate estimate of that parameter.

Rushikesh Kamalapurkar, Girish Chowdhary 0001, and Warren E. Dixon pioneered the use of historical data to solve this. In Concurrent learning for parameter estimation using dynamic state-derivative estimators, they describe a system that stores a history of past experiences—points where the system was excited—and uses them concurrently with current data.

By checking the current state against a "memory buffer" of diverse past states, the controller can maintain and even improve its parameter estimates even when the system is currently sitting still or moving in a straight line. This effectively replaces the need for future excitation with a record of past excitation.

Learning within Safety Boundaries

For safety-critical systems, the conflict is even sharper: the very noise required for learning might push the system into an unsafe state. Shengbo Wang and Shiping Wen address this by combining parameter estimation with Adaptive Control Barrier Functions (aCBF). This framework ensures that even if the system attempts to "probe" its environment for data, those probes are strictly filtered to keep the system within a predefined "safe set" Robust Finite-Time Adaptation Law for Safety-Critical Control.

Similarly, Aditya Gahlawat and Naira Hovakimyan developed $\mathcal{L}_1$-GP, which merges the rapid adaptation of $\mathcal{L}_1$ control with the non-parametric learning of Gaussian Processes L1-GP: L1 Adaptive Control with Bayesian Learning. This allows a system to learn complex, unknown dynamics in real-time without the "probing" signals that typically threaten stability.

Theoretical Necessity of Excitation

While engineers have sought to minimize PE for practical reasons, theorists like Paulo Tabuada and Lucas Fraile have questioned whether it was ever strictly necessary for stability. In A Note on Data-Driven Control for SISO Feedback Linearizable Systems Without Persistency of Excitation, they demonstrate that for a broad class of nonlinear systems, you can achieve stabilization using data-driven methods without satisfying the classical PE condition.

This suggests that the "wiggling" requirement was often an artifact of specific mathematical proofs rather than a fundamental physical law of control. By shifting to frameworks like Data-Enabled Predictive Control (DeePC) or reinforcement learning architectures that simulate experience Maopeng Ran, researchers are proving that autonomous systems can be both high-performing and "quiet," learning what they need to know from the natural flow of their mission rather than forced agitation.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_adaptive_control_without_persistent_exci_7f0aa6d0,
  title = {Adaptive Control without Persistent Excitation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adaptive-control-without-persistent-excitation/rdr_7f0aa6d083944a5894d496820957cf75},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_adaptive_control_without_persistent_exci_7f0aa6d0_v1,
  title = {Adaptive Control without Persistent Excitation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adaptive-control-without-persistent-excitation/rdr_7f0aa6d083944a5894d496820957cf75/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}