MIDGARD: A Robot Navigation Simulator for Outdoor Unstructured Environments

Giuseppe Vecchio, Riccardo Emanuele Sarpietro, Francesco Cancelliere, Simone Palazzo, Dario C. Guastella, Alessandro Strano, Ignacio Carlucho, Giovanni Muscato, Stefano V. Albrecht, Concetto Spampinato

Abstract: We present MIDGARD, a simulation platform based on Unreal Engine for training autonomous robots in complex outdoor unstructured environments. It offers photorealistic 3D scenes, procedural scene generation, and integration with ROS and OpenAI Gym. The focus of MIDGARD is on navigation, where an autonomous agent travels from random initial positions to designated target locations avoiding obstacles, enabling researchers to develop and evaluate novel algorithms and navigation methods. We evaluate MIDGARD’s suitability as a research tool by training navigation algorithms based on reinforcement learning; we also assess sim-to-real transfer capabilities in a traversable horizon prediction task, using deep learning models on RGB images only.