Corelab Seminar

Vasileios Tzoumas
Algorithmic Foundations of Trustworthy Collaborative Autonomy: From Robust Combinatorial Optimization to Perception and Control

Collaborative autonomous vehicles hold the promise to revolutionize transportation, disaster response, and space exploration. Already, micro-aerial vehicles with on-board cameras have become a multi-billion-dollar industry; and as we enter the new decade, teams of semi-autonomous flying cars, jet fighters, and space-exploration vehicles are being launched. An era of ubiquitous aerospace vehicles is becoming a reality, and along with it autonomous vehicles that can form teams, agree on navigation plans, and perceive the world. However, this future is threatened by denial-of-service (DoS) and deceptive attacks and failures that can compromise the vehicles' teams, navigation plans, and perception capabilities. These threats lie outside the reach of cybersecurity. Instead, algorithms at the intersection of perception, planning, and non-convex and combinatorial optimization are needed. I will present two algorithms from my research, and my vision for trustworthy collaborative autonomy.
First, I will discuss near-optimal algorithms for robust combinatorial optimization against any numbers of DoS attacks. The algorithms can robustify for the first time teams and their navigation plans against DoS attacks. I will demonstrate this via search and rescue, and surveillance experiments. Second, I will present algorithms that robustify visual perception capabilities against deceptive failures (outliers). The algorithms achieve extreme outlier-robustness in near real-time for the first time. I will illustrate this across various perception problems, on datasets for localization and mapping (SLAM), object recognition, and 3D-reconstruction. I will conclude with my vision for a collaborative autonomy that is not only robust but also resilient: I will argue the need for a technological convergence between (i) "cyber" capabilities for a distributed artificial intelligence, driven by adaptive learning and data-driven algorithms, and (ii) "physical" capabilities of self-reconfigurable structures.