Real-time Anticipation and Entrainment in Human-Robot Interaction

Abstract

In this work, reactive control methodologies, alongside real-time methodologies for dense human motion prediction, are utilized to facilitate real-time anticipation and entrainment in human robot interaction. The technical contributions of this thesis include: extensions to dynamical systems-based modulation approaches that enable real-time circumnavigation of non-convex obstacles (NOMAD), a trajectory clustering approach based on a relaxation of dynamic time warping (TRACER), a real-time human modelling and prediction approach (HABITS), and the integration of these technologies into an anticipation and entrainment controller that enables real-time adaptive synchronization between a human and a robot. NOMAD introduces several on-manifold strategies that enable real-time navigation in the presence of non-convex obstacles, alongside a methodology for the eff icient representation of dense environments that can represent up to 240k points while maintaining a 1ms loop. TRACER is a probabilistic trajectory clustering algorithm that uses the expectation-maximization algorithm and a relaxation of dynamic time warping (Soft-DTW), with demonstrable improvement over non-probabilistic techniques such as kMedoids or DBSCAN. HABITS is an event-driven probabilistic filtering and incremental profiling framework that provides robust segmentation, prediction, and alignment estimation in real-time (25Hz) for emergent interactions in structured settings. The combination of these technologies is then demonstrated to enable both effective real-time anticipation (de-conflicting a workspace), as well as to support entrainment (long-term human-robot synchronization) in human-robot interaction.