Online trajectory prediction and planning for social robot navigation

This paper addresses the safe and legible navigation of mobile robots in multi-agent encounters. A novel motion model provides the basis to predict, plan and coordinate agent trajectories in intersection scenarios. The approach establishes an implicit, non-overt cooperation between the robot and humans by linking the prediction and planning of agent trajectories within a unified representation in terms of timed elastic bands. The planning process maintains multiple topological alternatives to resolve the encounter in a manner compliant with the implicit rules and objectives of human proxemics. The trajectory is obtained by optimizing the timed elastic band considering multiple conflicting objectives such as fastest path and minimal spatial separation among agents but also global proxemic aspects such as motion coherence within a group. Cooperation is achieved by coupling predicted and planned agent trajectories to eventually reach an implicit agreement of the agents on how to circumnavigate each other. The parameters of the cost functions of the underlying motion model are identified by inverse optimal control from a dataset of 73 recorded encounters with up to five humans and a total of 283 individual trajectories. Playback simulations of recorded encounters and experiments with a robot traversing a group of oncoming humans demonstrate the feasibility of the approach to resolve general proxemic encounters.