Anticipatory Control on Human-Following Robots Using Online Deep-Model Predictive Control

Abstract

Mobile robots face challenges when collaborating with humans in crowded and occluded environments. To tackle this issue, we propose a solution called online deep model predictive control (Deep-MPC) and apply it to human-following robots. Deep-MPC incorporates a 3-D human detector, an online learning transition model, and a data-driven MPC framework. Specifically, the 3-D human detector generates the target's 3-D bounding box, while the transition model predicts future states, enabling anticipatory control. By combining the 3-D bounding box's intersection over union (IoU) and state anticipation, we propose a novel evaluation metric that enhances the following robustness. The data-driven MPC framework optimizes robot actions using the neural network of the transition model, and online learning occurs through autonomous interaction with the environment, eliminating the need for system modeling and controller design. To validate our method, we conducted extensive real-world human-following experiments, demonstrating its superior performance compared to some existing methods, skeleton-based methods, and approaches without Deep-MPC.