Control and Safety for Non-Affine Nonlinear Systems

This direction is merged into another direction: generalization to nonlinear systems. Its last version stays readable and citable here.

Version 1 of 185 papers87 conceptsfrozen 2026-10-04

In classical control theory, we often rely on a convenient mathematical fiction: the control-affine assumption. We assume the world behaves according to the equation $\dot{x} = f(x) + g(x)u$, where the control input $u$ enters the system linearly. This structure is the "skeleton key" of modern control; it allows us to easily invert dynamics, design Lyapunov functions, and guarantee safety.

However, as noted by Jingliang Duan and Shengbo Eben Li in Generalized Policy Iteration for Optimal Control in Continuous Time, real-world systems like vehicles with complex tire-road friction or chemical reactors often exhibit non-affine dynamics. In these systems, $u$ is buried inside complex nonlinearities—$\dot{x} = f(x, u)$—rendering standard tools for stability and safety ineffective.

Recent research has converged on a shared goal: developing a "universal" framework for controlling these complex systems by combining differential geometry, optimization, and machine learning.

The Stability Challenge: Beyond Lyapunov Functions

The traditional way to prove a system is stable is to find a Lyapunov function—a "bowl-shaped" energy function that always decreases. But finding these functions for non-affine systems is more of an art than a science.

Ian R. Manchester and Jean-Jacques E. Slotine proposed a shift in perspective called Control Contraction Metrics (CCMs) in Control Contraction Metrics and Universal Stabilizability. Instead of looking for a global energy function, contraction theory looks at the distance between any two neighboring trajectories. If all trajectories "contract" toward each other, the system is inherently stable.

The beauty of CCMs is that they transform the search for a controller into a convex optimization problem (Linear Matrix Inequalities), which is much easier to solve than the non-linear partial differential equations required by older methods. This approach provides a systematic way to stabilize systems that don't fit into rigid templates like "strict-feedback" forms.

The Safety Challenge: Barrier Functions in the Dark

For autonomous systems, stability isn't enough; they must also stay within "safe" boundaries. Control Barrier Functions (CBFs) have emerged as the standard tool for this, acting as a mathematical "force field" that prevents a system from entering unsafe states.

However, CBFs typically require a known, control-affine model. To bridge this gap, Jennifer Nguyen and David D. Fan introduced the BALSA framework in Bayesian Learning-Based Adaptive Control for Safety Critical Systems. BALSA uses Bayesian learning to estimate unknown dynamics while using stochastic CBFs to maintain safety.

Similarly, Hassan Almubarak and Evangelos A. Theodorou developed Gaussian Process Barrier States (GP-BaS) in Gaussian Process Barrier States for Safe Trajectory Optimization and Control. By using Gaussian Processes to model the uncertainty of non-affine dynamics, they can provide probabilistic safety guarantees even when the underlying physics are poorly understood.

Co-Synthesis: Learning the Proof and the Policy

One of the most powerful trends in this field is co-synthesis: using neural networks to learn the controller and the stability/safety certificate simultaneously.

In Learning Certified Control Using Contraction Metric, Dawei Sun and Chuchu Fan demonstrate how to train a neural network to act as a tracking controller while another network learns the CCM that proves its stability. This "closed-loop" learning ensures that the resulting controller isn't just high-performing in simulation, but has formal guarantees for real-world deployment.

For non-affine systems where the optimal control is hard to compute, Jingliang Duan and Qi Sun proposed Constrained Adaptive Dynamic Programming (CADP) in Adaptive dynamic programming for nonaffine nonlinear optimal control problem with state constraints. By locally linearizing the policy improvement step, they allow the system to respect state constraints even when the control input is embedded deep within a nonlinear function.

Solving the Hamilton-Jacobi-Bellman Equation

At the heart of optimal control lies the Hamilton-Jacobi-Bellman (HJB) equation, which is notoriously difficult to solve for high-dimensional, non-affine systems. Ziyi Wang and Evangelos Theodorou have pioneered a way to bypass this "curse of dimensionality" by transforming the HJB equation into Forward-Backward Stochastic Differential Equations (FBSDEs).

In Learning Deep Stochastic Optimal Control Policies using Forward-Backward SDEs, they show that by using LSTM neural networks to solve these SDEs, they can create scalable controllers for nonlinear systems that traditional numerical solvers cannot handle.

Summary of the Intuition

The shift from control-affine to non-affine control is essentially a move from algebraic convenience to geometric and data-driven robustness.

  1. Contraction (CCM) replaces global energy (Lyapunov) with local convergence, making the math tractable for general nonlinearities.
  2. Barrier Functions (CBF) are being "softened" with Gaussian Processes and Bayesian learning to handle model uncertainty.
  3. Neural Co-Synthesis allows us to learn the "proof" (the certificate) alongside the "action" (the controller), ensuring that deep learning doesn't come at the cost of safety.

By treating the controller and its stability certificate as a single, learnable unit, researchers are creating a new generation of robots that can navigate the complex, non-affine reality of the physical world with the same rigor once reserved for simple linear models.

Papers in this version

Showing 60 of 85 papers.

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_control_and_safety_for_non_affine_nonlin_b7a166e4,
  title = {Control and Safety for Non-Affine Nonlinear Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/control-and-safety-for-non-affine-nonlinear-systems/rdr_b7a166e434d64e5c92777a9a121dc5fe},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_control_and_safety_for_non_affine_nonlin_b7a166e4_v1,
  title = {Control and Safety for Non-Affine Nonlinear Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/control-and-safety-for-non-affine-nonlinear-systems/rdr_b7a166e434d64e5c92777a9a121dc5fe/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}