Adaptive Control for Nonlinearly Parameterized Systems

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 130 papers30 conceptsfrozen 2026-10-04

In classical control theory, we often assume that the world is "linearly parameterized." This means we assume that while we might not know a specific value—like the mass of a robot or the friction coefficient of a joint—that unknown value interacts with the system in a simple, linear way. Mathematically, we treat the uncertainty as a weight multiplied by a known feature.

However, real-world physics is rarely that cooperative. In complex environments like the open ocean, unknown parameters are often "baked into" the dynamics inside exponentials, trigonometric functions, or complex hydrodynamic drag equations. This is the problem of nonlinear parametrization. When parameters are buried this deeply, traditional adaptive algorithms fail to converge, and systems can become unstable.

Research in this area seeks to bridge the gap between rigid linear models and the chaotic, nonlinear reality of autonomous systems.

The Failure of the Linear Assumption

The core difficulty of nonlinear parametrization is that the relationship between what we observe and what we need to estimate is no longer straightforward. As Ivan Y Tyukin, Erik Steur, and colleagues point out in Adaptive Observers and Parameter Estimation for a Class of Systems Nonlinear in the Parameters, these systems are often "non-identifiable." Multiple different parameter combinations might produce the exact same observed behavior, making it impossible for a standard estimator to find the "truth."

In marine robotics, this isn't just a theoretical headache; it’s a barrier to autonomy. Yuanyuan Qiao and Wei Wang note in their Survey of Deep Learning for Autonomous Surface Vehicles that traditional hydrodynamic models struggle because wind, waves, and currents interact with a vessel's hull in ways that simple linear equations cannot capture. If a controller assumes a linear relationship for drag but encounters the exponential resistance of high-speed water flow, the vehicle will drift off course or oscillate wildly.

Geometric and Operator-Based Solutions

To handle these nonlinearities without losing mathematical rigor, researchers have turned to the geometry of motion. Instead of treating a vehicle as a point in flat space, Junwoo Jang and Maani Ghaffari propose using Lie Algebraic Model Predictive Control (MPC).

In Convex Geometric Trajectory Tracking using Lie Algebraic MPC, they show that by formulating motion on a Lie group ($SE(3)$), they can linearize the error dynamics rather than the system itself. This allows them to use fast, efficient solvers while respecting the nonlinear "shape" of the vehicle's rotation and translation. Ghaffari further extended this in Online Learning-Enhanced Lie Algebraic MPC by adding a lightweight Fourier-based learning module that adapts to environmental disturbances in real-time on low-power hardware.

Another powerful intuition comes from Koopman Operator Theory. In MAKO: Meta-Adaptive Koopman Operators, Minghao Han and Xunyuan Yin describe a way to "lift" nonlinear dynamics into a higher-dimensional space where they appear linear. This allows them to use the well-understood tools of linear adaptive control on systems that are fundamentally nonlinear, such as drones with shifting payloads.

Stability via Neural and Reinforcement Learning

While "black box" AI is often seen as the opposite of control theory, recent work integrates the two to provide safety guarantees. Ssu-Hsin Yu and Anuradha Annaswamy pioneered this in Neural Control for Nonlinear Dynamic Systems by using Lyapunov stability theory to constrain how a neural network learns. This ensures that even as the network adapts to nonlinearities, it won't push the system into an unstable state.

Modern approaches like Model-Reference Reinforcement Learning (MRRL), introduced by Qingrui Zhang and Wei Pan, use a simplified "nominal" model as a guide for a deep reinforcement learning agent. The agent learns to handle the complex, unmodeled nonlinearities (like wave slamming), while the nominal model provides a baseline of stability.

The cutting edge of this field even incorporates Large Language Models (LLMs). In the EasyUUV framework, Yubo Huang and colleagues use LLMs to "live-tune" controller parameters in the field. Similarly, Jingzehua Xu uses LLMs for high-level semantic adaptation in turbulent waters, as described in When Motion Learns to Listen.

Why It Works: The "Gentle" Adaptation

The shift in this research direction is away from "high-gain" control—where you simply overpower uncertainty with force—toward "gentle" or non-dominating algorithms. As Ivan Y Tyukin and Cees van Leeuwen argue in Adaptation and nonlinear parametrization, these non-dominating approaches are more energy-efficient and reduce wear on physical actuators.

By combining the structural insights of classical control (Lyapunov stability, Lie groups) with the flexible function approximation of modern AI (DRL, Koopman operators, LLMs), researchers are creating systems that don't just react to the environment, but internalize its nonlinear physics to navigate it more precisely.

Papers in this version

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_adaptive_control_for_nonlinearly_paramet_a1fc3368,
  title = {Adaptive Control for Nonlinearly Parameterized Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adaptive-control-for-nonlinearly-parameterized-systems/rdr_a1fc3368c47e43bd922aac340e0507cb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_adaptive_control_for_nonlinearly_paramet_a1fc3368_v1,
  title = {Adaptive Control for Nonlinearly Parameterized Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adaptive-control-for-nonlinearly-parameterized-systems/rdr_a1fc3368c47e43bd922aac340e0507cb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}