Task-aware waypoint sampling for robotic planning

Keren, Sarah and Canal, Gerard and Cashmore, Michael (2021) Task-aware waypoint sampling for robotic planning. In: Association for the Advancement of Artificial Intelligence Spring Conference Series, 2021-03-22 - 2021-03-24, Online. (https://drive.google.com/file/d/1GWMgHQD_oq0-VpCvt...)

[thumbnail of Keren-etal-AAAISCS-2021-Task-aware-waypoint-sampling-for-robotic-planning]
Preview
Text. Filename: Keren_etal_AAAISCS_2021_Task_aware_waypoint_sampling_for_robotic_planning.pdf
Final Published Version

Download (787kB)| Preview

Abstract

To achieve a complex task, a robot often needs to navigate in a physical space to complete activities in different locations. For example, it may need to inspect several structures, making multiple observations of each structure from different perspectives. Typically, the positions from which these activities can be performed are represented as waypoints – discrete positions that are sampled from the continuous physical space and used to find a task plan. Existing approaches to waypoint selection either iteratively consider the entire space or use domain knowledge to consider each activity separately. This can lead to task planning problems that are more complex than is necessary or to plans of compromised quality. Moreover, all previous approaches only consider geometric constraints that can be imposed on the waypoint selection process. We present Task-Aware Waypoint Sampling (TAWS), which offers two key novelties. First, it is an anytime approach that combines the benefits of random sampling with the use of domain knowledge in waypoint selection by performing a onetime computation of the connectivity graph from which waypoints are sampled. In addition, TAWS is the first approach that accounts for performance preferences, which are preferences a system operator may have about the generated task plan. These can account, for example, for areas near doorways where it is preferable that the robot does not stop to perform activities. We demonstrate the performance benefits of our approach on simulated automated manufacturing tasks.