Traditional model-based control strategies face a fundamental dilemma: they are only as good as the mathematical models underpinning them. In classical Nonlinear Dynamic Inversion (NDI) or Model Predictive Control (MPC), a controller relies on an explicit global physics model to calculate the exact actuator inputs required to achieve a desired state.
However, the real world is notoriously difficult to model. Micro Aerial Vehicles (MAVs) encounter unpredictable wind gusts, hybrid aircraft experience chaotic aerodynamic transitions, and robotic manipulators handle shifting payloads. Historically, researchers addressed these uncertainties by layering complex adaptive elements onto the system. For instance, Fuchun Sun 0001 et al. combined dynamic inversion with neuro-fuzzy systems to handle uncertain robotic manipulator dynamics without restricting state boundaries (Neuro-fuzzy adaptive control based on dynamic inversion for robotic manipulators). Similarly, Mohsen Soltani et al. designed a fallback mechanism using an internal model controller paired with motion sensors to maintain satellite tracking during external signal dropouts (Reliable Control of Ship-Mounted Satellite Tracking Antenna).
Yet, building a perfect global model remains an engineering bottleneck. Incremental Nonlinear Dynamic Inversion (INDI) bypasses this bottleneck entirely. Instead of attempting to model the entire environment, INDI uses high-frequency sensor feedback to measure the net effect of all forces acting on the vehicle, adjusting the control inputs incrementally to achieve the desired acceleration.
The Core Intuition: Why the Increment Eliminates the Model
To understand why INDI works, consider a generalized nonlinear system where the acceleration $\dot{x}$ depends on the current state $x$, external disturbances $d$ (such as wind), and the control input $u$:
$$\dot{x} = f(x, d) + g(x)u$$
In this formulation, $f(x, d)$ represents the complex, unmodeled aerodynamics and environmental disturbances—the hardest part to capture mathematically.
INDI resolves this by shifting the control objective from calculating an absolute control input $u$ to calculating an incremental change $\Delta u$ over a very small time step $\Delta t$. If the control loop operates at a high frequency (typically hundreds of Hertz), the physical state $x$ and the external disturbance $d$ change negligibly between consecutive steps. Therefore, the unmodeled dynamics at the current time step are approximately equal to those observed an instant ago:
$$f(x_t, d_t) \approx f(x_{t-\Delta t}, d_{t-\Delta t})$$
As detailed by Yingzhi Huang et al. in their application to complex hydraulic systems, applying a first-order Taylor expansion allows the system's behavior to be linearized locally around the current state (Incremental Nonlinear Dynamic Inversion Control for Hydraulic Hexapod Flight Simulator Motion Systems). By substituting the measured acceleration from the previous time step ($\dot{x}_{meas}$) into the expansion, the desired acceleration ($\dot{x}_{des}$) can be expressed as:
$$\dot{x}_{des} \approx \dot{x}_{meas} + B(x) \Delta u$$
Here, $B(x)$ is the control effectiveness matrix, which describes how a change in actuator position translates into a change in acceleration. Crucially, the complex, unmodeled term $f(x, d)$ has completely vanished from the equation. It is implicitly captured by the sensor measurement $\dot{x}_{meas}$.
Solving for the control increment yields:
$$\Delta u = B(x)^{-1} (\dot{x}_{des} - \dot{x}_{meas})$$
The final control command sent to the actuators is simply $u_t = u_{t-\Delta t} + \Delta u$. Because the controller focuses entirely on the difference between desired and measured acceleration, the onboard Inertial Measurement Unit (IMU) effectively acts as a proxy for sensing all unmodeled external forces, as demonstrated by Ewoud J. J. Smeur et al. (Cascaded Incremental Nonlinear Dynamic Inversion Control for MAV Disturbance Rejection).
Taming Nonlinear Aerodynamics in Flight Transitions
This mathematical simplification makes INDI highly effective for vehicles operating in volatile aerodynamic regimes. A prime example is the tailsitter UAV—a hybrid aircraft that takes off vertically like a helicopter and tilts 90 degrees to cruise horizontally like a conventional airplane. During the transition phase, the wings experience massive stall conditions and airflow separation, rendering global aerodynamic models useless.
Xinglu Xia et al. demonstrated that INDI provides exceptional robustness during these transitions precisely because it prioritizes observed state changes over global mathematical models (Transition Flight Control and Simulation of a Novel Tail-Sitter UAV With Varying Fuselage Shape). By focusing on local state increments, the controller maintains stability even when underlying aerodynamic coefficients are completely unknown.
Similarly, Ewoud J. J. Smeur et al. applied INDI to the Cyclone tailsitter, validating that high-speed sensor feedback allows the vehicle to absorb unmodeled aerodynamic forces in real time (Incremental control and guidance of hybrid aircraft applied to the Cyclone tailsitter UAV). By cascading two INDI loops—one for attitude and one for position—Ewoud J. J. Smeur et al. achieved significant improvements in wind gust rejection compared to traditional PID controllers, without requiring dedicated airspeed sensors (Cascaded incremental nonlinear dynamic inversion for MAV disturbance rejection). Further work by Ziqing Ma et al. on tilt-rotor tailsitters confirmed that tracking changes in state and actuator inputs enables smooth transitions without complex aerodynamic profiling (Design and Control of a Tilt-Rotor Tailsitter Aircraft With Pivoting VTOL Capability).
The High-Bandwidth Inner-Loop Shield
In modern autonomous control architectures, INDI is rarely used to plan long-term trajectories. Instead, it functions as a high-bandwidth inner-loop controller. Outer-loop planners generate optimal paths, while INDI acts as a robust shield, instantly canceling out high-frequency disturbances before they can destabilize the vehicle.
Sihao Sun et al. highlighted this division of labor in a comparative study of agile quadrotor flight, noting that a robust inner loop like INDI is the single most critical factor for real-world tracking accuracy when pushing a vehicle to its physical limits (A Comparative Study of Nonlinear MPC and Differential-Flatness-Based Control for Quadrotor Agile Flight).
This inner-loop isolation remains effective even when the state estimates themselves are noisy or unconventional:
- Vision-Based Control: Guido De Croon et al. demonstrated that INDI provides a robust framework for high-speed inner-loop control when relying entirely on optic flow, successfully absorbing the high variance and noise inherent in vision-based sensing (Accommodating unobservability to control flight attitude with optic flow).
- Bimodal Vehicles: Ruibin Zhang et al. utilized an INDI inner loop in terrestrial-aerial bimodal vehicles to seamlessly compensate for sudden changes in ground friction and wind during transitions between driving and flying (Model-Based Planning and Control for Terrestrial-Aerial Bimodal Vehicles with Passive Wheels).
Adapting to Structural Changes and Failures
Because INDI relies on the immediate physical response of the vehicle rather than a static model, it adapts dynamically when the vehicle's own physical configuration changes mid-flight.
Morphing and Asymmetric Payloads
When a drone changes its shape or carries an off-center load, its center of mass and inertia tensor shift instantly. Zhichao Han et al. leveraged an INDI controller to actively counteract the external torque disturbances and model uncertainties caused by a quadrotor morphing its shape to navigate confined spaces (Shape-Adaptive Planning and Control for a Deformable Quadrotor). Similarly, the authors of AutoTrans: A Complete Planning and Control Framework for Autonomous UAV Payload Transportation showed that an incremental control loop for angular velocity implicitly cancels out internal torques caused by hardware asymmetries or suspended payload swings without explicit modeling.
Actuator Failures
If a quadrotor experiences a sudden rotor failure, its dynamics change catastrophically. Sihao Sun et al. designed a fault-tolerant flight controller using nonlinear dynamic inversion that adapts its control allocation model on the fly (Autonomous Quadrotor Flight Despite Rotor Failure With Onboard Vision Sensors: Frames vs. Events). By re-allocating thrust increments based on the remaining three rotors, the system maintains stable position control without requiring an explicit model of the structural damage.
Actuator Saturation
When actuators are pushed to their physical limits, standard linear control allocation breaks down, causing severe oscillations. To address this, Alessandro Mancinelli et al. combined INDI with a real-time nonlinear control allocation framework using Sequential Quadratic Programming (SQP) (Real-Time Nonlinear Control Allocation Framework for Vehicles with Highly Nonlinear Effectors Subject to Saturation). This ensures that the calculated incremental changes ($\Delta u$) respect physical actuator saturation limits, allowing overactuated drones to remain stable in highly unpredictable environments.
Key Foundational Work
Reliable Control of Ship-Mounted Satellite Tracking Antenna (2010) — contributed 1 concept elements to this direction.
Neuro-fuzzy adaptive control based on dynamic inversion for robotic manipulators (2003) — contributed 1 concept elements to this direction.