Optimal Trajectories for Autonomous Human-Following Carts with Gesture-Based Contactless Positioning Suggestions

Abstract

Human-following autonomous robots can help human-oriented tasks in many fields. In this paper, we focus on replacing traditional shopping carts with human-following robots as a way to aid public health measures under events such as a global pandemic. The framework mirrors efforts made in other domains to introduce human-following robots into human-oriented tasks, and consists of two major building blocks: human pose estimation from 3D data, and collision-free navigation with gesture-based positioning suggestions. RGB-D data is used to estimate human poses and extract pointing gestures. The pose and pointing direction are used as positioning suggestions. The cart then either selects to move as close to the positioning suggestion as possible, or follows the human if they move too far away. The multi-cart system uses a model predictive control (MPC) based optimization method to generate multiple collision-free paths that satisfy as many positioning suggestions as possible. We validate our pathing method with both large-scale multi-cart simulations in ROS/Gazebo. We further demonstrate the human-following block with in-lab tests using a group of omni-directional robots.