ARRC Seminar Series - Dr. Gabriele Costante

Deep Learning (DL) and Deep Reinforcement Learning (DRL) methodologies have revolutionized a large number of scientific areas, and robotics is certainly among those that have benefited the most from them. In particular, DL and DRL have unlocked capabilities that were unthinkable just a few years ago, making it possible to practically deploy autonomous Micro Aerial Vehicles (MAVs) in several applications.

This seminar will outline current research and future directions on four topics that focus on the application of DL and DRL strategies to MAVs, specifically:

• Active Target Tracking. In Active Tracking an autonomous MAV has to track another dynamic object, such as another MAV. When the control commands to track the target object are computed from information gathered from vision sensors, the task is referred to as Visual Active Tracking (VAT).
• Vision-based Topological Localization. Topological localization has recently emerged as a promising alternative for robot localization. It aims to estimate the position of the robot with respect to a topological map, i.e., a sparse representation of the scene consisting of a set of nodes and a set of edges.
• Trajectory Tracking and Collision Avoidance. In many applications, MAVs are requested to follow a specific trajectory often computed on a simplified environmental model, which does not consider the actual scene structure and the presence of obstacles. Hence, MAVs have to track the reference trajectory as precisely as possible while avoiding collisions with obstacles that were not considered during planning. In practice, this implies that the MAV navigation system must be endowed with two core capabilities: Trajectory Following (TF) and Collision Avoidance (CA).
• Autonomous Exploration. Exploration is one of the most important tasks in robotics. While several researchers have addressed this problem, there are still many open challenges in the MAV contexts. In particular, the exploration trajectory might be significantly influenced by different factors, including limited flight time, safety constraints, and, most importantly, the objective of the exploration, e.g., 3D reconstruction, target search, detection of semantic entities. Therefore, classical path planning approaches might not provide optimal solutions and more advanced techniques are required.