Most physical systems—from quadrotors maneuvering through the air to chemical reactors processing materials—are inherently nonlinear. Controlling them directly requires solving complex nonlinear differential equations or running computationally expensive online optimizations. On the other hand, linear control theory is mature, computationally lightweight, and offers robust stability guarantees.
Feedback linearization bridges this gap. Instead of approximating a nonlinear system around a single operating point (like a traditional Taylor series approximation), feedback linearization uses an algebraic transformation to completely cancel out the system's nonlinearities. By injecting a clever nonlinear control law back into the system, the closed-loop system behaves exactly like a linear system from the perspective of an outer-loop controller.
As Frank Allgöwer noted in Chemical process control: a german perspective, exact linearization allows engineers to apply standard linear control tools to highly complex, non-proportional processes. The system is essentially "tricked" into behaving linearly, turning a daunting nonlinear control problem into a straightforward linear one.
The Mechanics of Linearization
To understand why this works, consider an input-affine nonlinear system of the form:
$$\dot{x} = f(x) + g(x)u$$
If we know the functions $f(x)$ and $g(x)$ precisely, we can choose a control input $u$ that neutralizes $f(x)$. For instance, setting $u = g(x)^{-1}(-f(x) + v)$ transforms the system into $\dot{x} = v$, where $v$ is a new virtual control input. The relationship between the virtual input $v$ and the state $x$ is now perfectly linear—a simple chain of integrators.
Depending on the system's structure and constraints, this strategy can take a few forms:
- Input-State Linearization: The entire state equation is transformed into a virtual linear system, an approach explored by Erkan Kayacan et al. in Experimental Validation of Linear and Nonlinear MPC on an Articulated Unmanned Ground Vehicle to achieve high-speed control.
- Input-Output Linearization: When the full state cannot be directly linearized, the control law focuses on creating a linear relationship between the inputs and specific chosen outputs. Helin Wang et al. leveraged this in Exponentially stable periodic walking of under-actuated biped robot to regulate the complex, coupled dynamics of underactuated walking robots.
- Nonlinear Dynamic Inversion: In aerospace applications, this technique cancels out complex aerodynamic terms. For instance, Takeshi Yamasaki demonstrated in Trajectory Guidance and Control for a Small UAV that inverting the dynamics transforms complex aerodynamic forces into a predictable linear system, enabling precise terrain following across varying flight speeds.
Differential Flatness and Trajectory Generation
A powerful extension of feedback linearization is differential flatness. A system is differentially flat if all of its states and inputs can be expressed entirely in terms of a specific set of carefully chosen outputs (called flat outputs) and their time derivatives.
This property acts as a mathematical shortcut. Instead of integrating complex nonlinear differential equations forward in time to plan a trajectory, we can plan a smooth path directly in the lower-dimensional "flat space" using simple polynomials or linear optimization. Once the flat trajectory is found, it can be mapped back into the original state and input space without losing any physical accuracy.
In DCAD: Decentralized Collision Avoidance With Dynamics Constraints for Agile Quadrotor Swarms, the authors utilized differential flatness to map non-linear quadrotor dynamics into a flat space, enabling high-speed trajectory generation with the computational efficiency of a linear solver. Similarly, Devansh R. Agrawal et al. showed in A Constructive Method for Designing Safe Multirate Controllers for Differentially-Flat Systems that this mapping allows complex robotic navigation tasks to be solved as real-time linear optimization problems.
Unlocking Computational Efficiency and Safety
The primary practical benefit of feedback linearization is computational speed. Nonlinear Model Predictive Control (NMPC) requires solving non-convex optimization problems at every time step, which can be too slow for high-speed or resource-constrained robots. By linearizing the system first, engineers can use standard Linear MPC (LMPC) or linear programming.
For example, Yasser Shoukry et al. in Scalable lazy SMT-based motion planning applied feedback linearization to scale motion planning up to 50 continuous states, bypassing the scaling limits of traditional nonlinear methods. In autonomous driving, Keenan Burnett et al. demonstrated in Building a Winning Self-Driving Car in Six Months that feedback-linearized controllers provide a robust, low-latency alternative to complex MPC under strict real-time update constraints.
Furthermore, feedback linearization simplifies formal safety verification. In Constructive Safety-Critical Control: Synthesizing Control Barrier Functions for Partially Feedback Linearizable Systems, M. Cohen et al. leveraged partial feedback linearization to systematically construct Control Barrier Functions (CBFs) for underactuated systems. By mapping the dynamics into a structured form, safety certificates can be derived more easily, as also shown by Stefan B. Liu et al. in Guarantees for Real Robotic Systems: Unifying Formal Controller Synthesis and Reachset-Conformant Identification, where feedback linearization enabled zonotope-based reachability analysis for nonlinear robots.
Handling Uncertainty and Real World Imperfections
The classic Achilles' heel of feedback linearization is its reliance on an accurate system model. If the mathematical model of $f(x)$ or $g(x)$ is incorrect, the cancellation will be imperfect, leaving residual nonlinearities that can destabilize the system.
Modern research addresses this by combining feedback linearization with learning and adaptation:
- Reinforcement Learning: In Feedback Linearization for Uncertain Systems via Reinforcement Learning, Tyler Westenbroek et al. proposed using model-free reinforcement learning to learn the linearizing control law directly from data, ensuring precise tracking even when the underlying physics are unknown.
- Adaptive Control and Model Standardization: Angela P. Schoellig et al. in Adaptive Model Predictive Control for High-Accuracy Trajectory Tracking in Changing Conditions introduced an adaptive layer that forces an uncertain system to match a predefined linear reference model, allowing an outer linear MPC to operate reliably despite external disturbances.
- System Identification: Thiago B. Burghi et al. in Feedback for nonlinear system identification turned the problem on its head, showing that a feedback-linearizing control loop can be used to isolate static nonlinearities from dynamic linear components, making the system easier to identify in the first place.
By transforming the geometry of the state space rather than ignoring its complexity, feedback linearization continues to serve as a foundational tool for making complex, high-dimensional systems manageable, safe, and computationally efficient.
Key Foundational Work
Chemical process control: a german perspective (1992) — contributed 1 concept elements to this direction.