Computer vision using machine learning in a Unity development environment

U radu se istražuje primjena i korištenje virtualnih okolina za treniranje agenata koristeći računalni vid. Cilj je stvoriti okolinu u kojoj će se agent kretati vizualnim opažanjima u svrhu obavljanja zadatka. Za stvaranje virtualne okoline korišten je Unity program za razvoj i upravljanje video igrama zajedno s ML-Agents paketom kao osnovom za treniranje agenta. Stvorena je okolina s tri stupca od kojih će jedan biti nasumično odabran kao cilj s metom na sebi dok su druga dva mamac agentu. Stupci su također generirani nasumične visine i pozicije. Agent je uspješno istreniran za prepoznavanje cilja i navigiranje prema njemu, izbjegavajući prepreke na putu.This work researches and applies usage of virtual environments in purpose of training agents using visual observations. It's purpose is to create an enviromnent that the agent will traverse using visual observations to complete his task. Unity game engine was used to create the environment alongside ML-Agents package that handles training the agent. The learning environment contained three pillars out of which one was randomly chosen to be the agent's goal with a visual target on top of it, whilst the other two were decoys. The pillars were also randomized in height and position. The agent was successfully trained to spot his target as well as to navigate towards it while avoiding obstacles