Robust Following with Hidden Information in Travel Partners.

Pedestrians often travel with one another [9]. Travel partners share destinations or temporary subgoals. Such information is often communicated on-the-fly, when observations are received and local direction decisions are needed. At decision points, e.g. intersections, we observe that there is often a leader (or more) in a group who actively decides where to go, and the follower(s), without explicit communication, can adapt their motions quickly to catch up [10]. In robotics, human-following is achieved by performing online subgoal inference, for the robot to adapt the motion while maintaining desired group shapes [5], such as side-by-side walking. This capability has been realized by applying maximum-likelihood subgoal estimates upon the robot path planning problem [8, 10]. While assertively following towards the most likely subgoal has shown success in relatively simple environments, such strategy can lead to bad states under false inference, e.g. poor visibility to other route options or blocking group member path options, yielding poor performance under inference delay.