Online Learning and Planning for Crowd-aware Service Robot Navigation

Abstract

Mobile service robots are increasingly used in indoor environments (e.g., shopping malls or museums) among large crowds of people. To efficiently navigate in these environments, such a robot should be able to exhibit a variety of behaviors. It should avoid crowded areas, and not oppose the flow of the crowd. It should be able to identify and avoid specific crowds that result in additional delays (e.g., children in a particular area might slow down the robot). and to seek out a crowd if its task requires it to interact with as many people as possible. These behaviors require the ability to learn and model crowd behavior in an environment. Earlier work used a dataset of paths navigated by people to solve this problem. That approach is expensive, risks privacy violations, and can become outdated as the environment evolves. To overcome these drawbacks, this thesis proposes a new approach where the robot learns models of crowd behavior online and relies only on local onboard sensors. This work develops and tests multiple planners that leverage these models in simulated environments and demonstrate statistically significant improvements in performance. The work reported here is applicable not only to navigation to target locations, but also to a variety of other services.