Abstract:

At CLeAR, we believe human models are crucial for fluent human-robot collaboration. When coupled with planning/learning, human models enable robots to display intelligent interactive behavior. However, crafting good human models remains difficult. In this talk, we’ll discuss facets of human modeling, starting with our efforts in capturing human trust in the robot. We’ll discuss a perspective that views trust as function learning and the impact this has in multi-task scenarios. Next, we will cover MIRROR, a new approach to quickly learn human models that are useful for (communication) planning. MIRROR is inspired by social projection theory, which hypothesizes that humans use self-models to understand others. In a similar manner, MIRROR leverages robot self-models learned using RL to bootstrap human modeling. We will examine how MIRROR performs relative to state-of-the-art methods using simulations and a human-subjects study. Finally, we’ll discuss how MIRROR can be extended towards real-time human-robot collaboration, and take a broader perspective of the challenges and opportunities that lie ahead. 

>> If you have any questions, please don't hesitate to contact me,