Active Perception for Robot Teams: From Visual Search to Videography, Micah Corah
Sprache des Titels:
Englisch
Original Kurzfassung:
Over the last ten years, drones have become increasingly integrated with our society: drones film our sports, inspect our crops, survey our geography, and inspect our disaster sites. Across these domains, drones are key because they are particularly adept at maneuvering cameras and sensors to ideal vantage points in diverse environments. However, these applications often still involve either manual operation or extensive operator interaction, and the teams deploying these systems can consist of multiple operators per robot. Bridging this gap will require both more effective coordination between robots and better understanding of application domains.
My work focuses on enabling aerial robots to make intelligent decisions about how to sense, sample, and observe their environments both individually and in groups. I will start this talk by discussing active perception with individual robots in the context of searching for survivors in a subterranean environment; I will discuss how robots can quickly navigate and map such environments with careful attention to dynamics, camera views, and the interactions between the two. Given individual robots that we have endowed with the ability to intelligently observe and inspect, how can we develop teams that coordinate effectively and efficiently? Toward this end, I will turn to the problem of autonomously filming a group of people such as to film a team sport or a dramatic performance. By applying the rich theory of submodular and combinatorial optimization, simple algorithms can enable individual robots that are able to film autonomously and augment them with the ability to coordinate in teams. I will then present a distributed submodular optimization algorithm I developed (Randomized Sequential Partitions or RSP) that enables this approach to scale to large numbers of robots, and I will discuss how to apply this approach to multi-robot videography by carefully designing objectives and reasoning in terms of pixel densities.
Forschungs-