The control of unmanned aircraft systems must be rigorously tested and verified to ensure their correct functioning and airworthiness. Incorporating components that use novel techniques such as deep learning can pose a significant challenge because traditional approaches for detecting errors require deriving a model of a correctly performing controller, which can be intractable. An alternative approach is to specify conditions that have to be satisfied by such controllers and assess the behavior of the system using a technique that has become known as adaptive stress testing. This simulation-based approach provides a way to find the most likely path to a failure event, accelerating the search by formulating stress testing as a sequential decision process and then optimizing it using reinforcement learning.

Leading and following can emerge naturally in human teams. However, such roles are usually predefined in human-robot teams due to the difficulty of scalably learning and adapting to each agent’s roles. Our goal is to enable a robot to learn how leader-follower dynamics emerge and evolve in human teams and to leverage this understanding to influence the team to achieve a goal. We focus on developing algorithms for robot leaders that can influence human teams. We plan to develop an effective, concise, and scalable team structure representation by combining model-based methods and data-driven techniques to uncover the underlying human team latent structures. We aim to optimize for a robot policy that leverages the leader-follower graph to attain an objective by influencing human teams.

ACAS sXu is a protocol for collision avoidance for small drones.  In this project we aim to do a formal robustness analysis of a specific implementation of sXu being used by GE.  This implementation uses a deep neural network (DNN).  The goal is to show that the advisory produced by the DNN for each point in the testing set is sufficiently robust (i.e. no nearby point produces a different advisory) to ensure that there are no untested points that could produce an advisory that differs dramatically from the tested points.  The verification effort includes using and improving as needed our neural network verification tool Marabou (based on the Reluplex calculus).

Adaptive stress testing (AST) is an approach for validating safety-critical autonomous control systems such as self driving cars and unmanned aircraft. The approach involves searching for the most likely failure scenarios according to some measure of likelihood of occurrence. To enhance efficiency when searching the enormous space of possible scenarios, AST applies reinforcement learning to a black-box simulation of the autonomous agent and its environment. Current research on AST is focused on the development of new reinforcement learning algorithms and objective functions to improve efficiency and performance. Additional research is focused on incorporating constraints on the failure modes that AST finds (e.g., failures that occur in one control system but not another or failures that were the explicit fault of the autonomous agent), as well as expanding fault coverage.