Área: Engenharia Informática e de Computadores

Despacho de nomeação de Júri

Título da Tese: Ad Hoc Teamwork in the Real-World

Local da Prova:  Anfiteatro PA-3 (Piso -1 do Pavilhão de Matemática) do IST

Data: 18/07/2025

Hora: 14h00
Abstract: This thesis studies the problem of Ad Hoc Teamwork, where autonomous agents are deployed on-the-fly to unexpected situations and required to assist existing teams without pre-coordination or pre-communication. Specifically, it explores the narrower setting of Ad Hoc Teamwork in Real-World Domains, where environments may be partially observable, actions of the teams not visible to the agent, and teams hybrid, composed not only by other robots but also humans as teammates. It asks the research question of "How can an autonomous agent assist teammates in performing tasks under more realistic conditions, without being able to pre-coordinate or pre-communicate with the existing team?", which it breaks down into four individual sub-problems, each addressing its own challenge: - How can an agent efficiently assist teammates and tasks on-the-fly, without having to learn how to interact with them from zero? - How can an agent assist teammates in performing tasks without being able to observe their actions, the full state of the environment or both? - How can an agent assist teammates in performing tasks in arbitrarily large, partially observable domains? - How can an agent assist mixed human-robot teams in performing tasks in real-world scenarios? Each sub-problem is respectively addressed with the proposal of the following contributions: - TEAMSTER, a model-based reinforcement learning approach, which allows an agent to continuously adapt to teammates and tasks by efficiently reusing environmental knowledge. - BOPA and ATPO, two Bayesian online approaches capable of identifying and assisting teammates in performing tasks without needing to observe their actions. - ATLPO, an approach capable of identifying and assisting unknown teammates in performing unknown tasks in arbitrarily large partially observable domains without needing to observe their actions, relying on state-of-the-art deep learning techniques. - The HOTSPOT framework, an ad hoc teamwork platform for mixed human-robot teams that enables the deployment of ad hoc approaches into real-world scenarios.