Search

Browse

• Communities

My Space

• Login

Control of a self-balancing two-wheeled vehicle using computational intelligence methodologies

Cheiladakis Dimitrios

Full record

---

Summary

Balancing a two-wheeled vehicle is a problem that has drawn attention from both researchers and industry. These systems are inherently unstable—if no corrective action is applied, they tip over almost immediately. Because of this, they are often compared to the well-known inverted pendulum problem, a benchmark in control engineering. Beyond being a theoretical challenge, however, technology has very practical uses. It underpins personal mobility devices such as the Segway, appears in service and warehouse robots, and is being explored in assistive devices that can enhance mobility for people with physical limitations. This combination of practical relevance and theoretical difficulty makes the study of self-balancing vehicles particularly appealing.A striking feature of such vehicles is how directly they reveal the effectiveness of a control strategy. When the controller works, the vehicle remains upright; when it fails, the vehicle quickly falls. This immediacy makes them not only good case studies for advanced control research but also valuable educational tools. At the same time, it highlights the difficulty of the problem: controllers must process real-time sensor inputs—often from accelerometers and gyroscopes—while compensating for noise, drift, and delays. These real-world issues mean that approaches which perform well in theory do not always succeed in practice.The dynamics of a self-balancing vehicle are nonlinear and highly sensitive to disturbance. A small irregularity in the ground, a push from the side, or a shift in load can destabilize it. Controllers must therefore adjust continuously, updating the torque applied by the motors in fractions of a second. Classical approaches such as PID control, state feedback methods, and Linear Quadratic Regulators (LQR) have all been applied successfully to this challenge. However, these approaches tend to rely on accurate mathematical models of the system. In reality, models are never perfect, and performance can deteriorate once uncertainties and nonlinear effects come into play.Another practical issue with classical controllers is the difficulty of parameter tuning. For instance, a PID controller requires carefully chosen gains, but values that work well under one condition may fail when conditions change. Similarly, LQR designs depend on weight matrices that reflect trade-offs between competing goals such as stability and energy use. In practice, these values are often found through trial and error, which is time-consuming and sometimes unreliable. This dependence on tuning motivates the search for alternatives that can cope with variation more naturally.One promising direction is the use of computational intelligence techniques. Approaches such as fuzzy logic control, neural networks, and neuro-fuzzy systems are attractive because they do not require a precise model of the system. Fuzzy logic makes it possible to describe control rules in intuitive, human-like terms—for example, “if the tilt is small but increasing quickly, apply a strong correction.” Neural networks, in contrast, excel at learning from data and can capture complex nonlinear relationships that are hard to model directly. A hybrid neuro-fuzzy system can combine these advantages, offering both interpretability and adaptability.The strength of these approaches lies in their robustness and flexibility. Unlike conventional controllers, they can adapt to unexpected conditions and cope with unmodeled dynamics, sensor noise, or changes in the environment. For a self-balancing vehicle, this can translate into better stability on rough terrain, under varying loads, or when disturbances are applied. Beyond this particular application, similar methods have been successfully used in domains such as autonomous driving, robotics, and unmanned aerial vehicles, showing that the underlying ideas are widely applicable.This thesis investigates these methods in the context of a self-balancing two-wheeled vehicle. Its primary goal is to design, implement, and compare controllers based on both classical control and computational intelligence. A lab-scale prototype built with the Arduino Engineering Kit Rev 2 and the Arduino Nano 33 IoT is used as the experimental platform. This makes it possible to test different approaches under controlled conditions and evaluate their relative strengths and weaknesses. The research aims to answer a central question: can computational intelligence methods provide a practical advantage over traditional controllers in stabilizing self-balancing vehicles?

Available Files

Services