Interactive decision-making for autonomous vehicles based on a theory of order

Abstract

Autonomous driving is challenging, because understanding how human-being makes decisions can be difficult. To build such a machine that substitutes for human to make decisions, one essential is the capability of interacting with other road users.Currently autonomous vehicles lack such an understanding, thus over conservative behaviors must be taken, which leads to compromised performances. The subject of this thesis is the development of the Theory of Order and the order-based approaches as the tool, to understand human behavior, hence enable autonomous vehicles to make interactive decisions. The theory is a mathematical framework for describing the strategies a human-like mind would take during the traffic interactions modelled as a game. It aims to cut the issue of multi-equilibria and incomplete information that makes the game difficult to solve. The meaning of Order is threefold: it describes a set of steady states that generally exists in the interactions, as a subjection relation of behaviors among the players, which is not precisely known but could be dynamically approximated through a process of order establishment; it also indicates the sequence in which the traffic space is occupied by players, and how objective and random differences would lead to the inevitability of the final subjection relation; and it's a strategy profile regarding player's aggressiveness that establish a high level equilibrium in the sense of considering driving as infinite repeated game, therefore provides a solution to multi-equilibria.The order establishment could be achieved by players adjusting their intentions constantly according to the order-strategy, which is explored both on a theoretical and on an algorithmic level. By conducting various simulations in both structured space (unsignalized intersection) and unstructured space (Arc de Triomphe roundabout), the approach is tested and verified in completing specific tasks, including unprotected left-turn, merging, or free interaction in proximity. Compared to the baseline methods, order-based approach shows its advantages in several aspects: promoting both individual and system efficiency, generating collision free maneuver, being capable of taking initiative actions in a safe manner. Exhaustive test reveals that system level performance would be promoted if players are relatively xiii xiv equivalent in aggressiveness, which indicates that AV should be designed adaptive when dealing with different groups of people. In general, this work propose an innovative framework to interpret and analyse traffic interactions and developed algorithm that improve performances in safety and efficiency for the autonomous vehicles to interact with humans.