Simultaneous Prediction and Planning in Crowds using Learnt Models of Social Response

Abstract

The ability of autonomous mobile robots to work alongside humans and animals in real world environments has the potential to revolutionise the way in which many routine and labour intensive tasks are completed. Whilst we are seeing increasing applications in controlled environments, such as traffic and warehousing, robots are still far from ubiquitous in everyday life. In unstructured environments, such as agriculture or pedestrian crowds, where interactions between agents are not guided by infrastructure, there exist additional challenges that need to be overcome before we are likely to see the widespread adoption of mobile robots.

Safe navigation in shared environments requires the accurate perception of nearby individuals using a robot's on board sensors. Additionally, the future motion of detected individuals needs to be predicted both for collision avoidance and efficient navigation. These predictions should reflect the inherent uncertainty of the individual's future, including the ways in which an individual might respond to its neighbours, including the robot itself. As such, there exists a dependency between any prediction of an individual's motion and the planned path of the robot, which needs to be accounted for both during the prediction and planning stages of navigation.

This thesis focuses on how prediction and planning can be approached in a single framework to address this dependency, using learnt models of social response within a sampling based path planner for simultaneous prediction and planning (SPP). Additional challenges faced in navigating shared and unstructured environments are also addressed, including predicting the uncertain branching and multi-modal nature of agent motion during social interactions, and overcoming the on-board limitations of mobile robots --- such as resource and sensing constraints --- in order to achieve extended autonomy.