In this work, we present an algorithm for autonomously determining the appropriate location from which to observe a human or robot agent (actor) while it completes a task in dynamic environments. We develop theory for selecting such a location using forward physical simulation of randomly-selected candidate viewpoints. The simulated points provide obstacle avoidance, and by incorporating a modified version of the Social Force Model, candidate viewpoints adjust themselves so that they do not encroach on the actor's personal space and/or safety region. The best observer position is chosen from these candidates to provide the most complete view of the task volume, taking into account the occlusion caused by the actor itself. We show that our algorithm works under a variety of task volume configurations, actor types (human and robot), and environmental constraints. Finally, the paper shows the results of hardware deployment on a two-robot system - one observer, and one actor. The paper concludes by examining the social impacts of deploying autonomous observation algorithms on real-world systems.
We have made our code available on GitHub. Please use the GitHub issue tracker for any bug reports.
If you have any other questions or comments, or to get in touch with the authors, please email allevato (at) utexas (dot) edu.