Recognition of operator motions for real-time assistance using virtual fixtures

Abstract

Hidden Markov Models (HMMs) are used for automatic segmentation and recognition of user motions. A new algorithm for real-time HMM recognition was developed. The segmentation results are used to provide appropriate assistance in a combined curve following and object avoidance task. This assistance takes the form of a virtual fixture, whose compliance can be altered online. Recognition and assistance experiments were performed using force and position data recorded from a cooperative manipulation system, where a robot and a human operator hold an instrument simultaneously. Recognition accuracy exceeds 90%, even when the users training the HMMs differ from those executing the task. For a task consisting of both path following and avoidance motions, an HMM-based virtual fixture switches the compliance from low to high when the user is trying to move away from the path. The HMM method improves operator performance in comparison with a constant virtual fixture and no virtual fixture.