Tuning Cost Functions for Social Navigation

Human-Robot Interaction literature frequently uses Gaussian distributions within navigation costmaps to model proxemic constraints around humans. While it has proven to be effective in several cases, this approach is often hard to tune to get the desired behavior, often because of unforeseen interactions between different elements in the costmap. There is, as far as we are aware, no general strategy in the literature for how to predictably use this approach. In this paper, we describe how the parameters for the soft constraints can affect the robot's planned paths, and what constraints on the parameters can be introduced in order to achieve certain behaviors. In particular, we show the complex interactions between the Gaussian's parameters and elements of the path planning algorithms, and how undesirable behavior can result from configurations exceeding certain ratios. There properties are explored using mathematical models of the paths and two sets of tests: the first using simulated costmaps, and the second using live data in conjunction with the ROS Navigation algorithms.