Learning to Recognize Discrete State Transitions in Assembly 1

Assembly tasks can be characterized by the discrete states of contact between the parts being assembled. However, the robot sensory and control systems operate in the continuous domain, leading to the problem of associating states and events of the physical system with states and events in the discrete task space. This paper presents a new approach to discretizing sensory data, based on discriminant functions and clustering techniques , for applications in robotic process monitoring and in interpreting human sensory data. The advantage of using linear discriminant functions for discretizing the sensory data is that these functions are learned/trained using supervised learning. This enables the process monitoring system to improve its performance with practice and to track changes in the assembly process parameters over the long term. Clustering techniques are used to classify the sensory data for use by discriminant function learning algorithms. Since clustering is a form of unsupervised learning, the combination of clustering techniques and linear discriminant functions provides the process monitoring system with the ability to automatically interpret continuous sensory data without any human intervention. The method presented here was applied to sensory data obtained from a force sensor during an assembly task and was found to successfully discretize the data.