An autonomous guided vehicle can be used for the delivery of goods. To deliver these goods to your home, the mobile robot will be driving in pedestrian rich environments. In these environments the robot will need to socially navigate itself in a way which is both pleasant for the pedestrians and progressive for the robot. During its drive, the robots behavior should be adapted to the situation it finds itself in. The underlying theory for this is the Social, Technology and Service triangle. This triangle dictates the balance between three aspects. The technology aspect focuses on power consumption and efficiency. The service aspect is all about delivery time and location. The social aspect is all about minimizing social disruption and facilitating intuitive behavior. The behavior of pedestrians can be modelled with the Social Force Model. In this model the objects and pedestrians apply social forces onto each other which determine their movements. The robot can use this SFM for its own social navigation. Multi-Policy Decision Making (MPDM) can be added onto the Social Force Model. This allows for switching between three basic policies: go-solo, follow and stop. Through forward simulations the robot can predict and decide which of the three policies brings the most progress and the least social disruption. However, SFM-MPDM on its own does not cope well with passing and crossing; thus methods were designed to handle these situations. In passing the robot looks ahead and makes room, often keeping right. In crossing the robot slows down and deviates to cross behind the pedestrian. The Social Force Model, the Multi-Policy Decision Making and the passing and crossing all combined form the social navigational model. This model consist of the many different parameters which govern the behavior and is part of the overall computational mechanism. This mechanism takes in the environment through sensors such as LiDAR. It preprocesses these inputs and sends them to the model. The output of this model is what determines the locomotion and behavior of the robot. To find the behavior and its underlying parameters an Evolutionary Algorithm was employed.
The learnt behavioral parameter sets were tested with participants (n=42) and a physical ROSbot. They were asked to rate the robots behavioral performance on comfort, predictability and communication of intent. From these results, it became clear that the robots behavior influences the experience of the pedestrians, but it is unclear which parameter exactly influences which part of the behavior. There are certain expectations to what impact a parameter or group of parameters has, but at which times and to what effect this controls parts of the behavior is difficult to identify. Nevertheless a good foundation has been laid down for future projects through a social navigational model and a way to learn parameters for the many different situations an autonomous delivery robot can encounter. A recommendation is to improve the Evolutionary Algorithm in order to facilitate parameter learning which better matches the desired behavior in the STS-triangle.