October 28, 2024
Title: Multi-task learning for rapid online adaptation under Signal Temporal Logic specifications in autonomous systems
Location: Walker Engineering 320
MS Teams link: Join virtually (Meeting ID: 261 842 829 456; Passcode: x9hZ3Z)
Affiliation: Department of Aerospace Engineering, Mississippi State University
Abstract: In this session, a Multi-Task Learning (MTL)-based control framework that considers Signal Temporal Logic (STL) specifications is presented. The main goal is to improve the generalizability of the controller in new tasks, exploiting useful information incorporated in related tasks. MTL settings involve learning and testing stages. In the learning stage, an ensemble of tasks is generated by perturbing STL specifications. Task compliance is measured via the robustness degree, which is computed using the STL semantics. In the testing stage, new unseen tasks are generated. The solution from multitask learning is used as a warm-start, leading to fast, few-shot adaption to new tasks. Both stages, learning and testing are solved using Sequential Convex Programming to deal with the non-convex nature in the robustness degree expression. The methodology is applied to the system dynamic of a quadcopter within a scaled framework of the Air Traffic Control problem.