Data-Driven Intention Inference and Its Application to Human-Robot Coordination

As robots become increasingly integrated into daily life, their ability to effectively coordinate with human-maneuvered agents, including humans, cars, and drones, becomes crucial. However, existing techniques for human-robot coordination have treated all human agents uniformly, without considering the unique characteristics of each agent. This PhD dissertation proposes a research framework that incorporates the specific attributes of individuals involved in the coordination process and aims to design adaptive coordination abilities for robots.

Inspired by how humans interact and collaborate with familiar individuals, this research investigates how robots can learn from human behavior to enhance their coordination capabilities. Humans intuitively infer each other’s intentions and take into account the historical behavior of others to improve collaboration. The proposed research tackles three key aspects: 1) Developing an efficient approach to model human intentions based on historical behavior data. Different types of intentions are classified, and each intention is mapped to a corresponding motion pattern. 2) Inferring human intentions and utilizing the model to determine the current motion pattern for effective coordination. 3) Incorporating inferred intentions and motion patterns into robotic applications. The proposed approaches are applied to two specific applications: state estimation and robotic escorting.

The effectiveness of the proposed approaches is validated through simulations and real experiments. The novel model for intention and motion patterns demonstrates significant advantages in efficiently describing human behavior. In both the state estimation of a human-maneuvered quadrotor and the robotic escorting of a human using a mobile robot, the proposed approaches exhibit benefits such as higher estimation accuracy, enhanced flexibility, and improved user satisfaction.