Dynamic fleet management for autonomous vehicles

Background

The rise of autonomous vehicles is a present-day topic. Different leading companies are rapidly developing automated cars, deriving benefit from the new possibilities in information and communication technology. The application potential of autonomous vehicles on existing logistical problems is enormous, and a huge impact on society is expected.

Due to increasing transportation demands and highly specific requirements of customers, current logistics problems become more and more complicated. Combining passengers or loads into one single vehicle and reducing the number of empty trips can decrease the congestion and pollution levels significantly. Furthermore, adequately handling real-time information and using stochastic information to deal with uncertainty can improve the efficiency and reduce the costs of an operating fleet. To obtain this, coordination and cooperation between vehicles is important, which will be possible for autonomous vehicles due to recent ICT developments and individual computing power. However, optimizing the fleet behaviour in real time is a very hard problem.

Goal

Although the Vehicle Routing Problem and its extensions have been studied for years, its models and solutions are still limited. The main research challenges in the field of autonomous fleet management are in developing rich models that reflect many real-world properties and developing methods for large scale problems. Hence, the question we focus on in this research is:

How can a fleet of autonomous vehicles be used for efficient and robust handling of a large number of customer-specific transportation tasks in a dynamic and uncertain world?

In particular, we focus on problems where transportation orders can have multiple pickup and/or delivery locations and/or time windows, to deal with the problem of customers not being at home during delivery. If they specify for example being at home in the morning and at work in the afternoon, the transport operator has more flexibility, and less unattended deliveries can be obtained.

Approach

In addition to existing exact and (meta)heuristic methods, we focus on distributed approaches since they are likely to be useful for dynamic, urgent, and large-scale problems. We consider a multi-agent approach in which order agents are each responsible for one transport order, and vehicle agents for the trajectory of one vehicle. By local interactions and negotiations between the agents, a solution has to be found. Due to the local and parallellizable computations, this approach can provide reasonable solutions where centralized approaches that search for optimal solutions are insufficient. Furthermore, advantages of flexibility, reactivity, and robustness are obtained.

Figure: Routing solution for a problem with multiple pickup and delivery locations per customer and preferences for the different options.

Contact:

PhD candidate: Johan Los
Daily supervisor: Prof. dr. R.R. Negenborn/Dr. M.T.J. Spaan
Email: J.Los@tudelft.nl