**1.2 Classification**

CAs differ in terms of weight, size, kind, altitude, payload, and a variety of other characteristics. According to their type and height, they can be divided into two broad categories (**Figure 1**). Both category have their own set of benefits and drawbacks. Various sorts of CAs are utilised depending on the application scenario. **Table 1** shows the classification of CAs.

Classification based on the weight of UA (Unmanned Aircraft) as follows:


#### **Figure 1.**

*Classification of CAs on the basis of the design.*



#### **Table 2.**

*Technical parameters of CAs.*


Typical physical parameters of small CAs for commercial applications can be summarised as follows as in **Table 2**.

## **2. State of art in CA**

CA, sometimes known as drones, has had robust growth in the previous 5 years all over the world. The model UAS fleet is expected to grow from 1.25 million entity to about 1.39 million by 2023, according to the study aerospace projection fiscal years 2019–2039, while the non-model CA fleet is expected to rise from 277,000 CA to over 835,000 CA by 2023. CA's beneficial applications have the potentiality for saving lives, improving safety and efficiency, and allow for more impactful engineering as well as research [7]. Designers experimenting with small CA for a variety of purposes such as aerial surveillance as well as personal recreational flying, entrepreneurs exploring parcel and medical supply delivery, and search and rescue missions are just a few examples.

While CA have their origins in military uses, they have recently become more helpful towards scientific and commercial purposes [8, 9]. Remote sensing, georeferencing, cartography, customs and border protection, investigation, rescue operations, fire espial, agronomic imaging, traffic surveillance systems, and package delivery are just a few of the applications they have recently discovered around the world.

Due to the rapid growth of CA technology, the extensive usefulness of CAs for numerous applications has been recognised, ranging from transportation services to disaster search and rescue.

While many current control systems still rely heavily on the availability of precise mathematical models (e.g., Model-Predictive-Control (MPC) [10], linear quadratic Gaussian [11, 12], backstepping [13], as well as gain scheduling [14]), this article evaluates extra versatile and intelligent approaches by emphasising the value of evolutionary computation to resolve the actual constraints of model-based control systems.

When building a robust flight control system, there are a few things to keep in mind. The first issue is the closed-loop control's robustness in the presence of uncertainties [11], including unpredictably extremely high air passes (e.g., violent wind gusts) and modelling errors. A small CA's mobility can be extremely vulnerable to wind gusts, which might cause the system to deviate from its intended trajectories. This phenomenon can also result in large overshoots and tracking offsets, both of which are undesirable in terms of safety and efficiency.

While many current dynamic control systems still rely significantly on mathematical equations of the subsystems (e.g., gains scheduling [14] as well as feedback linearization strategies), these approaches may be excessively complex or unworkable in some cases. Gain scheduling control, for example, has been considered one of the most historically dominant adaptive control approaches, but it has a number of technical flaws. Because it significantly leans on the linearization technique of the aviation dynamics over numerous places in the performance envelope, as well as several joint interpolation approaches, the system is extremely mathematical and timeconsuming. It could potentially result in a system that lacks global property. Furthermore, in the absence of thorough mathematical models, feedback linearization could be impractical.

Despite the positive results, MBC-based designs face a hurdle in that they rely on the correctness of the mathematical model of a real plant. According to imprecise system information and omnipresent exogenous disruptions, a poorly developed or described model might have a negative impact on later controller synthesis, resulting in inadequate performances or even instability. Uncertainties and disturbances of this nature can be categorised as follows:


To address these issues, a variety of modern control systems have been offered, each with its own set of benefits, restrictions, and drawbacks. Gain Scheduling (GS) [10], for example, is a frequently used strategy that shows good capabilities in dealing with parametric variations and nonlinearities, but frequent and fast changes in the controller gains might make the system unstable [13]. Furthermore, as noted in [14], the cost of implementation rises with the frequency of functioning points. Robust control, on the other hand, is effective when dealing with constrained parametric uncertainties, but it has drawbacks when dealing with boundless ones or stochastic dynamics [17, 18]. Adaptive control is a potential method for managing parametric uncertainties (because to its real-time adaptation strength); nonetheless, there are few commonly acknowledged approaches here to robust adaptive control issue so far [19]. The sliding control technique has been demonstrated to be resistant to modelling mistakes and parameter uncertainty, however frequent controller switches can cause chattering. Furthermore, when exogenous disruptions occur, the insensitivity to parameter changes characteristics may cause problems with self-stabilisation. Last but not least, thanks to using a continually updated model, namely an ultra-local model, Model-Free Control (MFC) approaches that have arisen to tackle stochastic dynamic behaviour as well as ambiguities of nonlinear systems have exhibited outstanding adaptation and estimating capabilities. However, for the time being, this methodology is confined to system dynamics that can be turned into Single-Input Single-Output (SISO) subsystem. There are other issues with analytic stability and evidence of convergence. ANNs have been used to analyse complicated control systems in order to solve the previously mentioned limitations of MBC-based solutions. This is primarily due to ANNs' perceived advantages in structural analysis and controller design [14, 20], which include their ability to recognise stochastic and multinomial systems [21, 22], their capacity to adapt in real-time, and their relatively simple computation methodology and hardware implementation.

As a result of these characteristics, ANNs are a fantastic tool for building the systems underneath prototype of high accuracy and low sophistication, even if it is distorted by uncertainties and disturbances, as well as for facilitating the implementation process and improving real-time performance. Regardless, there are still obstacles owing to their data-driven essence that limit their industrial applications, to some extent, due to various: a need for huge datasets of training data; the tendency to learn spurious relationships, which can lead to poor generalisation functionalities [23]; dearth of readability due to their own black-box characteristics [24]; and the lack of a structured method for pertaining ANN architecture designs [25] (In other words, given a certain ANN design, the number of hidden layers and synapses, the sort of perceptron, weight update algorithms, and so on are often decided haphazardly than in a structured manner).
