Using sensor fusion and contextual information to perform event detection during a phase-based manipulation task

In this paper we present an approach to event detection during a dextrous manipulation task. The approach utilizes a combination of tactile sensors as well as contextual information. The manipulation task is decomposed into distinct phases, each of which is associated with a limited number of feasible events such as making or breaking contact, slipping, etc. A set of context-based and sensor-based features is associated with each possible event for each type of manipulation phase. The goal is to detect events as reliably and as rapidly as possible. At any time during a task, each possible event is assigned a confidence value between 0 and 1. This indicates how confident the detection scheme is that a given event could be occurring at that instant. A high-level controller can then make use of this information to determine when to switch to a different manipulation phase.