Hierarchical Dynamic Games for Human-Robot Interaction with Applications to Autonomous Vehicles

Abstract: The deployment of autonomous vehicles has become an active area of research due to its potential solution as a more efficient and intelligent transportation system. While early research focused on isolated systems, autonomous vehicles integrated on public roads have lately received substantial attention. Here, the autonomous vehicle needs to interact with other drivers, in particular human drivers. Accurate interaction models are then crucial: the actions of one driver will affect the actions of other drivers sharing the road, and vice versa. This thesis presents a human-robot interaction model for autonomous vehicles with focus on longer-time planning. More precisely, the interaction between an autonomous vehicle and a human driver is captured as a hierarchical game. The hierarchical game predicts future actions of the two agents by partitioning its planning horizon into two sub-horizons: one high-fidelity horizon modelling immediate interactions and one low-fidelity horizon approximating remaining longer-time interactions. The objectives of the agents are captured via reward functions and the solution of the game is obtained as a probabilistic variant of the feedback Stackelberg solution. The solution is used by the autonomous vehicle to plan its optimal control trajectory, typically in a receding horizon manner. The model is validated via simulations and case studies. The former involve two cars acting according to the hierarchical game. Behaviour such as overtaking and merging emerges naturally as a consequence of the game. The case studies consider instead an autonomous truck platoon and a human driver, where the hierarchical game can be integrated into the platoon controller by restricting the human interaction to one truck at a time. The resulting platoon opens up a wider truck gap next to the human if the human is subject to danger in its lane (the human can then merge into the platoon lane avoiding the danger), while discouraging such cut-ins otherwise by keeping a tight formation. This behaviour is again not hard-coded but emerges from the game. Both the simulations and the case studies are compared with shorter look-ahead alternatives. Our model yields more realistic behaviour and better performance.