Deep reinforcement learning based obstacle evasion

Postavljena je simulacija u programskom sustavu Unity. Uporabljene su tri različite scene za učenje. U sceni se nalazi dron koji se može gibati u dvije dimenzije, s fiksiranom visinom. Naučen je model arhitekture Pix2Pix za mono-okularnu predikciju dubine.. Korišteni su podaci generirani u okviru simulacije. Algoritam izbjegavanja prepreka je Q-učenje. Simulacija daje algoritmu RGB sliku prostora ispred drona. Predviđa se dubina i zajedno sa željenim smjerom daje se kao ulaz u duboku Q-mrežu (DQN). Model se sastoji od potpuno povezanih i konvolucijskih slojeva. Ostvarena je i varijanta s LSTM slojem. Radi usporedbe su korištene i neke heuristike za izbjegavanje prepreka. Najbolji rezultat je postignut kombinacijom predviđene dubine i varijantom algoritma polja potencijala. A simulation is built using Unity. There are three different testing scenes. A drone is placed in the scene. It can only move in two dimensions. The height is fixed. Pix2Pix architecture is used for single camera depth prediction. Training dataset is generated from the simulation. Algorithm used for obstacle evasion is Q-learning. Simulation provides algorithm with RGB image of the area in front of the drone. Predicted depth with target angle is passed as input to a Deep Q-Network (DQN). Model is composed of dense and convolution layers. There is also a DQN version which uses a LSTM layer. Several obstacle avoidance heuristics are developed to better compare results of the DQN. The best result is achieved by using predicted depth with potential field algorithm.