People Find-and-Follow Behavior for Service Robots using Continuous Belief Maps
Abstract
Find-and-follow is an important behavior for social robots which assist people. We introduce an on-line planning method for the robot: Continuous Real-time POMCP (CR-POMCP). It uses Partially Observable Monte-Carlo Planning (POMCP), which, in contrast to most other planning algorithms, can plan under uncertainty, on large state spaces, and real-time. Our method, CR-POMCP, also works in continuous state spaces, and moreover takes into account sensory noise, false negatives, and false positives.
Comparisons have been done with a Heuristic Follower, which simply follows the person. Two extensions to CR-POMCP were tested to improve the robot's performance when doing real-life experiments. All variants tested in simulation were found to be working better than the Heuristic Follower. Dynamic obstacles were introduced in simulation by adding randomly walking people. Real-life experiments have been done during a week with a mobile service robot in three urban environments of Barcelona with other people walking around.
Experiments
The robot has to search and follow the person. The person to follow is recognized using AR Markers.
Heuristic Follower vs. Adaptive HB-CR-POMCP Follower
The Heuristic Follower goes to the last position where it saw the person, and when it reached that position it waits until it sees the person. In contrast, the Adaptive HB-CR-POMCP Follower has a belief (i.e. 'memory') with the location of the person.
The video shows two experiments on the FME lab, where first the Heuristic Follower does not move until the person is visible. The Adaptive HB-CR-POMCP Follower, in contrast, does move, trying to find the person actively.
Adaptive HB-CR-POMCP Follower in the BCN Lab
The method is shown to work in different large urban environments. The robot finds and follows the person.