Authors

Yurii Vakula, Viktoriia Kravtsova, Nataliia Miniailo, Siddhartha Pundit

Abstract

The project investigates an experimental, AI-based approach to tuning proportional–integral–derivative (PID) controllers for high-degree-of-freedom robotic platforms, beginning with acustom mini quadcopter. Rather than relying only on manual adjustment or a fully constructedphysics model, it uses behavior and telemetry observed from the real device to search forsuitable coefficients with a genetic algorithm. The immediate scope is three PID coefficientsapplied to the two horizontal axes; broader, higher-dimensional tuning is planned for laterstages.

The initial physical platform is a custom ESP32 mini quadcopter mounted on a three-axis testbench. The bench provides a controlled starting point for experiments that could becomeunstable, while the onboard flight controller retains final authority. The team framed theproblem around platforms whose motor thrusts, balance, centre of mass and externaldisturbances make simplified coefficient tables less adequate than for steady systems. Abalance vector can first compensate for construction and motor differences, giving PID tuninga near-equilibrium starting state.

The proposed workflow begins with experimental identification tests on the real platform.Telemetry is used to form a delta model relating control effort to changes in angle, angularrate, acceleration and velocity. The model’s observed limits constrain a fitness function, afterwhich the genetic algorithm evaluates candidate coefficient sets. Lower fitness values aretreated as better. Selected candidates are then intended to return to the bench for physicalvalidation, with repeated cycles potentially producing a map of settings for different conditionsand disturbances.

At system level, a workstation communicates bidirectionally with the quadcopter through a Wi-Fi gateway. The demonstrated stream contained measured states and motor-control efforts,while commands can be sent in the reverse direction. Tuning is a one-time procedure for eachdevice, but the framework is intended for many individual platforms, each treated according toits own characteristics. Safety is addressed through the constrained bench and an onboardangle limit that stops a test and can shut down the device.

license

Creative Commons BY-SA 4.0