In this work, we introduce MIDGARD, an open-source platform designed for training autonomous robots
to
navigate in outdoor unstructured environments. MIDGARD is tailored to facilitate the development and
deployment of autonomous agents, such as unmanned ground vehicles, in photorealistic 3D settings that
closely mimic real-world conditions.
Key features of MIDGARD include a flexible and extensible procedural landscape generation pipeline that
can adjust difficulty levels, coupled with rapid scene rendering capabilities powered by Unreal Engine.
Furthermore, MIDGARD offers a Python programming interface that allows for the integration of new sensor
types and customization of internal simulation variables.
The platform also provides a range of simulated agent sensors, including RGB, LiDAR, depth, and
instance/semantic segmentation, making it an ideal tool for developing and benchmarking diverse navigation
solutions.
We validate the effectiveness of MIDGARD through on-field tests of a navigation approach trained on the
platform, demonstrating its impressive sim-to-real capabilities. These results underscore MIDGARD's
potential as a robust simulation environment for autonomous navigation.
@article{vecchio2022midgard,
title={MIDGARD: A simulation platform for autonomous navigation in unstructured environments},
author={Vecchio, Giuseppe and Palazzo, Simone and Guastella, Dario C and Carlucho, Ignacio and Albrecht, Stefano V and Muscato, Giovanni and Spampinato, Concetto},
journal={arXiv preprint arXiv:2205.08389},
year={2022}
}