Lecturer: Jangho Choi (University of California)

Title: Data-Driven Frameworks for Safety Verification and Applications in Advanced Air Mobility

Date & Venue: November 13 (Thursday), 4:00–5:30 PM, L409

Abstract:
Decision-making in aviation involves many complex layers—from operational-level decisions like route scheduling, to trajectory-level decisions such as air traffic control, to vehicle-level decisions like flight stabilization. In this talk, I will focus on emerging safety challenges across these layers and explore how tools from machine learning and control theory can help address them.
Specifically, I will introduce two projects that combine control-theoretic tools for safety guarantees with data-driven approaches for ensuring safety at both the vehicle dynamics level and the air traffic control level.

First, recent aviation incidents have highlighted how understaffed air traffic management (ATM) systems can impact safety, underscoring the urgent need for automation. While learning-enabled tools like multi-agent reinforcement learning (MARL) excel at strategic decision-making in diverse scenarios, their lack of formal safety guarantees limits their use in aviation.
I will introduce a safety challenge in multi-agent scenarios—known as the “leaky corner” issue—and demonstrate how we can resolve it to achieve Safe MARL for automated ATM by integrating MARL with safety certificate functions from control theory.

Next, recent advances in electric propulsion technologies have led to the emergence of novel vertical take-off and landing (VTOL) aircraft designs, enabling real-world applications like air-taxi services. However, these new designs introduce safety challenges at the vehicle control level due to uncertain and nonlinear dynamics, especially during flight mode transitions.
I will present the concept of a Data-Driven Hamiltonian, which integrates data with control objectives to provide safety guarantees. This approach enables the construction of dynamically feasible safe sets of the VTOL aircraft directly from the vehicle trajectory data.