Efficient and Trustworthy Artificial Intelligence for Critical Robotic Systems
Time: Tue 2022-11-29 14.00
Location: F3, Lindstedtsvägen 26 & 28, Stockholm
Subject area: Computer Science
Doctoral student: Christopher Sprague , Robotik, perception och lärande, RPL
Opponent: Associate professor Todor Stoyanov, McMaster University, Hamilton, Ontario, Canada
Supervisor: Professor Petter Ögren, Robotik, perception och lärande, RPL; Associate Professor John Folkesson, Robotik, perception och lärande, RPL; Ivan Stenius, Farkostteknik och Solidmekanik
Critical robotic systems are systems whose functioning is critical to both ensuring the accomplishment of a given mission and preventing the endangerment of life and the surrounding environment. These critical aspects can be formally captured by convergence, in the sense that the system's state goes to a desired region of the statespace, and safety, in the sense that the system's state avoids unsafe regions of the statespace. Data-driven control policies, found through e.g. imitation learning or reinforcement learning, can outperform model-based methods in achieving convergence and safety efficiently; however, they often only do so by encouraging them, thus, they can be difficult to trust. Model-based control policies, on the other hand, are often well-suited to admitting formal guarantees of convergence and safety, thus they are often easier to trust. The main question asked in this thesis is: how can we compose data-driven and model-based control policies together to encourage efficiency while, at the same time, formally guaranteeing convergence and safety?
We answer this question with behaviour trees, a framework to represent hybrid control systems in a modular way. We present the first formal definition of behaviour trees as a hybrid system and present the conditions under which the execution of any behaviour tree as a hybrid control system will formally guarantee convergence and safety. Moreover, we present the conditions under which such formal guarantees can be maintained when including unguaranteed data-driven control policies, such as those coming from imitation learning or reinforcement learning. We also present an approach to synthesise such data-driven control policies in such a way that they encourage convergence and safety by adapting to unforeseen events. Alongside the above, we also explore an ancillary aspect of robot autonomy by improving the efficiency of simultaneous localisation and mapping through imitation learning. Lastly, we validate the advantages of behaviour trees' modularity in a real-world autonomous underwater vehicle's control system, and argue that this modularity contributes to efficiency, in terms of ease of use, and trust, in terms of facilitating human understanding.