When controlling a physical system—whether it is an autonomous underwater vehicle navigating ocean currents, a high-precision robotic arm fighting joint friction, or a quadrotor buffeted by wind—engineers face a fundamental challenge: the gap between idealized mathematical models and the messy reality of the physical world.
Traditional control paradigms rely heavily on accurate models. If the model is wrong, or if the environment introduces unexpected forces, the controller's performance degrades. To counter this, engineers historically spent immense effort deriving complex first-principles models for every physical nuance, from aerodynamic drag to mechanical wear.
Extended State Observation (ESO) flips this paradigm on its head. Instead of trying to model every complex, time-varying internal uncertainty and external force individually, the ESO lumps them together into a single, unmeasured virtual state—the "total disturbance"—and estimates it in real time using the system's actual input-output behavior. By estimating this aggregate disturbance, the controller can actively cancel it out, transforming a chaotic, unpredictable system into a well-behaved, predictable one.
The Core Insight: Lumped Disturbance and State Augmentation
To understand why this works, consider a standard second-order dynamic system, such as a moving mass:
$$\ddot{x} = g(x, \dot{x}) + d(t) + b u$$
Here, $x$ is the position, $u$ is the control input (force), $b$ is the control gain, $g(x, \dot{x})$ represents internal dynamics (like friction or gravity), and $d(t)$ represents external environmental disturbances (like wind gusts).
In practice, we rarely know $g(x, \dot{x})$ or $d(t)$ perfectly. The core mathematical trick of the ESO framework is to rewrite this equation by defining a new, unified variable $f(x, \dot{x}, d, t)$, known as the total disturbance or lumped uncertainty:
$$f(x, \dot{x}, d, t) = g(x, \dot{x}) + d(t) + (b - b_0)u$$
where $b_0$ is our rough estimate of the control gain. The system equation simplifies beautifully to:
$$\ddot{x} = f + b_0 u$$
As highlighted by Ben M. Chen et al. in their overview of complex intelligent control systems, this formulation treats all unknown internal dynamics and external environmental forces as a single, unified disturbance term (Complex system and intelligent control: theories and applications).
The clever leap is to treat this total disturbance $f$ as an additional, hidden state variable ($x_3 = f$). We then augment our standard state-space model (which normally only tracks position $x_1$ and velocity $x_2$) to include this virtual third state:
$$\begin{aligned} \dot{x}_1 &= x_2 \\ \dot{x}_2 &= x_3 + b_0 u \\ \dot{x}_3 &= \dot{f} \end{aligned}$$
By framing the unknown physics as a state variable, we can use state estimation techniques to track it dynamically. As noted in the development of stabilized gyro platforms, this strategy of unified disturbance lumping simplifies complex system dynamics by treating friction, mechanical clearances, and parameter fluctuations as a single observable state, bypassing the need to model each physical phenomenon individually.
How the Extended State Observer Tracks the Invisible
How can an observer estimate $x_3$ (the disturbance) if we don't have a mathematical formula for $\dot{f}$?
The observer exploits a simple physical truth: any discrepancy between how we expect the system to accelerate (based on our control input $u$) and how it actually accelerates (measured via sensors) must be caused by the total disturbance $f$.
A linear Extended State Observer mimics the augmented system structure but adds correction terms based on the estimation error:
$$\begin{aligned} \dot{\hat{x}}_1 &= \hat{x}_2 + \beta_1 (x_1 - \hat{x}_1) \\ \dot{\hat{x}}_2 &= \hat{x}_3 + b_0 u + \beta_2 (x_1 - \hat{x}_1) \\ \dot{\hat{x}}_3 &= \beta_3 (x_1 - \hat{x}_1) \end{aligned}$$
Here, $\hat{x}_1, \hat{x}_2, \hat{x}_3$ are the estimated position, velocity, and total disturbance, while $\beta_1, \beta_2, \beta_3$ are observer gains.
Think of it as a closed-loop tracking system inside the controller's software. If the real system moves faster than the observer predicts ($x_1 - \hat{x}_1 > 0$), the error term pumps a correction signal directly into the disturbance estimate $\hat{x}_3$ via the gain $\beta_3$. As long as the observer's internal dynamics are tuned to be faster than the rate of change of the physical disturbance, $\hat{x}_3$ will rapidly converge to the true value of $f$.
As Chaodong Jiang et al. demonstrate in high-precision motor systems, this allows the observer to capture time-varying lumped disturbances and dynamic uncertainties that are entirely absent from static physical models (Friction-Compensation and Extended State Observer Based Model Predictive Control for PMSM Servo Systems).
Active Disturbance Rejection: Achieving the Ideal Plant
Once the observer provides an accurate, real-time estimate of the total disturbance ($\hat{x}_3 \approx f$), the controller can execute Active Disturbance Rejection Control (ADRC).
Instead of relying on slow, reactive feedback loops (like a standard PID controller that only responds after an error has accumulated), we can use a proactive feedforward law. We design our control input $u$ to actively cancel out the estimated disturbance:
$$u = \frac{u_0 - \hat{x}_3}{b_0}$$
If we substitute this control law back into our simplified system equation ($\ddot{x} = f + b_0 u$), we get:
$$\ddot{x} = f + b_0 \left( \frac{u_0 - \hat{x}_3}{b_0} \right) \approx u_0$$
The real-world, non-linear, disturbed system has been mathematically stripped of its uncertainties. From the perspective of the virtual control input $u_0$, the plant now behaves like an ideal, disturbance-free chain of integrators ($\ddot{x} = u_0$).
u_0 ---> [ + ] ---> [ / b_0 ] ---> ( Real Plant: ) ---> Output (x)
^ \ddot{x} = f + b_0*u |
| |
[ - ] <--- \hat{x}_3 (Disturbance) |
| |
[ Extended State ] <------------------------------+
[ Observer ] <--- Control Input (u)
This architecture decouples reference tracking from disturbance rejection. As explained by Hongjia Li et al., the ADRC framework integrates this real-time estimation and compensation loop to neutralize both internal and external perturbations before they can degrade tracking performance (Fuzzy Optimized MFAC Based on ADRC in AUV Heading Control).
Advanced Frontiers: Speed, Structure, and Safety
While the basic linear ESO is highly effective, modern applications push the boundaries of how quickly and accurately these observers can track disturbances.
Finite-Time and Chattering-Free Estimation
Traditional linear observers guarantee that estimation errors decrease asymptotically (approaching zero as time goes to infinity). However, for high-speed systems like underwater vehicles or quadrotors, the observer must catch up to the disturbance before the system drifts off course.
To achieve faster convergence, researchers integrate non-linear sliding-mode techniques into the observer structure. For instance, Long He et al. utilize a super-twisting extended state observer (STESO) to track complex aerodynamic and hydrodynamic forces (Three-Dimensional Trajectory Tracking Control of Underactuated AUV Based on Fractional-Order PID and Super-Twisting Extended State Observer). By embedding a super-twisting algorithm into the observer correction terms, the system achieves finite-time convergence and eliminates the high-frequency oscillations ("chattering") that plague traditional sliding-mode observers.
Handling Unmatched Disturbances
The standard ADRC framework assumes that the disturbance enters the system through the exact same "channel" as the control input (known as a matched disturbance), allowing for direct algebraic cancellation. But what happens if a disturbance affects a state that the control input cannot directly touch?
To solve this, Xinming Wang et al. developed high-order disturbance observers designed for safety-critical systems (Safety-Critical Disturbance Rejection Control of Nonlinear Systems With Unmatched Disturbances). These advanced observers estimate not only the magnitude of the external forces but also their higher-order derivatives. This allows the controller to mathematically map how an unmatched disturbance propagates through the system states over time, enabling proactive compensation to prevent safety violations before they manifest in the primary output.
By treating the unknown world not as a modeling failure, but simply as an unmeasured dimension of the system state, Extended State Observation provides a robust, elegant bridge between clean control theory and the unpredictable realities of physical deployment.