Human integration into discrete event control systems is considered by utilising potential fields for control synthesis. Human integration into a control system is of importance in many situations, including those in which limited sensor information about the environment or the task is available. The framework presented allows the sharing of human commands with commands from an automated control system. Potential fields are being used as a tool to generate velocity commands from an autonomous task level controller as well as allowing the human to interact. The potential fields are also utilised to constrain human input such that human input error is minimised. The ideas presented here are supported by experiments.

No abstract

A process monitor for robotic assembly based on hidden Markov models (HMMs) is presented. The HMM process monitor is based on the dynamic force/torque signals arising from interaction between the workpiece and the environment. The HMMs represent a stochastic, knowledge-based system in which the models are trained off-line with the Baum-Welch reestimation algorithm. The assembly task is modeled as a discrete event dynamic system in which a discrete event is defined as a change in contact state between the workpiece and the environment. Our method (1) allows for dynamic motions of the workpiece, (2) accounts for sensor noise andfriction, and (3) exploits the fact that the amount of force information is large when there is a sudden change of discrete state in robotic assembly. After the HMMs have been trained, the authors use them on-line in a 2D experimental setup to recognize discrete events as they occur. Successful event recognition with an accuracy as high as 97% was achieved in 0.5-0.6 s with a training set size of only 20 examples for each discrete event.

This paper presents a model-based approach to the recognition of discrete state transitions for robotic assembly. Sensor signals, in particular, force and moment, are interpreted with reference to the physical model of an assembly process in order to recognize the state of assembly in real time. Assembly is a dynamic as well as a geometric process. Here, the model-based approach is applied to the unique problems of the dynamics generated by geometric interactions in an assembly process. First, a new method for the modeling of the assembly process is presented. In contrast to the traditional quasi-static treatment of assembly, the new method incorporates the dynamic nature of the process to highlight the discrete changes of state, e.g., gain and loss of contact. Second, a qualitative recognition method is developed to understand a time series of force signals. The qualitative technique allows for quick identification of the change of state because dynamic modelling provides much richer and more copious information than the traditional quasi-static modeling. A network representation is used to compactly present the modelling state transition information. Lastly, experimental results are given to demonstrate the recognition method. Successful transition recognition was accomplished in a very short period of time: 7-10 ms.

A new technique for selecting, in real time, different sensing techniques for a robotic system has been developed. The proposed method is based on stochastic dynamic programming, which provides an effective solution to multi-stage decision problems. A t each stage in the decision process a sensor selection controller has the option of consulting a new process monitoring technique to improve the knowledge of the task or terminating the decision process without any further information gathering. The sensor selection controller has been successfully implemented for the real-time control of a planar robotic assembly task in a discrete event control framework. One of the monitoring methods used is based on Hidden Markov Models, where the average recognition rate was 87%. Larger recognition rates for the HMM method have been demonstrated by 141. The rate of 87% was chosen to show the efectiveness of the dynamic sensor selection method. The experiments show that the method performs better than any individual process monitor. A successful event recognition rate of 97% with an average CPU time of 0.38 seconds is achieved when two force monitors and one position monitor are available to the sensor selection controller.