Semantic environment representation has gained a lot of attention in the robotics community in the last few years. The ability for a robot to understand its environment in a "human-like" way through NN-based object detection and segmentation can significantly extend and streamline autonomous navigation and exploration behavior, for example detecting dangerous areas, doors/openings, or ground terrain. Lately, this commonly visual detection process has been extended into using 3D pointclouds from onboard LiDAR sensors instead of RGB camera images, through the PointNet1 or similar software.

The Robotics and AI group has worked in subterranean navigation for Autonomous Unmanned Aerial Vehicles for many years in collaboration with the mining industry, but recently also in relation to the search for life in subterranean systems on other planets through a collaboration with NASA’s Jet Propulsion Laboratory. The team has developed a robust autonomy framework, denoted as COMPRA2 (COMPact Reactive Autonomy) that has seen deployment in many realistic scenarios3. COMPRA generates very consistent navigation behavior, but is limited in its ability to handle complex environments.

In addition to a general investigation of the PointNet (or similar) for robotic systems, this master thesis can also include an extension of the COMPRA framework through a semantic environment representation such as but not limited to: junction detection, tunnel/opening detection, dead-end detection, or detection of objects of interest, as to link an abstract environment representation to the navigation behavior of the robot through pointcloud-based object detection. This master thesis is for a student with an interest in Artificial Intelligence, Neural Nets, and programming, and for someone who has an inquisitive drive to test novel state-of-the-art systems in real applications, and who can work towards a real-life implementation of their developed framework for a robotic system.

• The student will investigate state-of-the-art Neural Nets for Pointcloud segmentation and object detection in order to classify critical environment parameters.

• The aim is the utilization of such frameworks for subterranean or indoor areas where the desired robot navigation behavior greatly depends on what kind of environment it is in.

• The work can, in a straight-forward way, be integrated to the COMPRA framework and evaluated in real-life autonomous mission scenarios.

• The work will preferably be integrated into the Robot Operating System (ROS).