Kaiji Itabashi

Realization of skill controllers for manipulation of deformable objects based on hybrid automata

The requirement of handling deformable objects such as leather, paper and rubber is growing. Since it is very difficult to make a physical model of them, the design of a controller to manipulate them becomes one of the significant problems in the field of robotics. If we look at the operation of a human worker, the deformable objects seem to be handled naturally and smoothly. The paper presents a method to design a controller for assembly tasks which involves the manipulation of deformable objects. The proposed method extracts the dynamics that human workers used in executing the demonstrated task, and embeds that in an event driven hybrid controller. In our control system, an event observer estimates the change of task state based on force and visual information like human workers, and switches the dynamics appropriately according to the task state. The proposed method is applied to a hose insertion task by implementing it in an industrial robot controller.

Modeling of the peg-in-hole task based on impedance parameters and HMM

When we apply an impedance control to execute any tasks, it is very important how to decide the impedance parameters to realize the desired task. If we can extract impedance parameters from human teaching data as characteristics of the human skill, it is appropriate to use them for control because of the similarity between an impedance control and a human fingertips control. However, there often exists unevennesses in time and space in human data. Modeling with hidden Markov model (HMM) is one of the promising technique to construct an efficient model for time-variant data including unevennesses. HMM is capable of characterizing a doubly stochastic process with an underlying immeasurable stochastic process which can be measured through another set of stochastic processes. Therefore, the probabilistic modeling of certain time series data which includes unevennesses caused by the human is possible. In this paper, we propose a method to model the series of impedance parameters identified from human teaching data with HMM in order to extract an essential discrete model which expresses the human skill. In addition, some applications of the obtained model to robot control and skill evaluation are shown.

Quantitative evaluation for skill controller based on comparison with human demonstration

One of the promising strategies to design a skill controller for robots is to observe the human workers skill and embed it in the robot controller under certain control architecture. However, no systematic design strategies to realize this scenario have yet been developed due to the lack of a quantitative performance evaluation of the skill controller. In this brief, the switching-impedance controller is considered as the skill controller and is developed based on a comparison with human workers demonstration. The enabling condition to switch the impedance parameter is optimized by calculating a hidden Markov model (HMM) distance which can measure the similarity between the skill of the human worker and the robot. HMM is a doubly stochastic system and is recognized as a useful tool for speech recognition. Thanks to the similarity in the stochastic characteristics between speech and skill (position/force) data, HMM is also expected to play a crucial role in skill controller design. An insertion task of deformable objects with the assistance of a vision sensor is considered in this brief. Some parameters which appear in the skill controller are optimized so as to increase the similarity with human workers demonstration.

Modelling and realization of the peg-in-hole task based on hidden Markov model

Impedance control is widely used in the field of industrial world. In a certain task, it is important to decide the impedance parameters in order to realize the desired task. However, it is very difficult to calculate analytically, and the method to extract impedance parameters from human demonstration often exist unevenness in time and space in the human data. Modelling with hidden Markov model (HMM) is known as one of the promising technique to construct an efficient model for time-variant data including unevenness. HMM is capable of characterizing a doubly stochastic process with an underlying immeasurable stochastic process which can be measured through another set of stochastic processes. In this paper, we propose a method to model the series of impedance parameters identified from human teaching data with HMM as human skill model of the peg-in-hole task. In addition, realization method of the task based on the obtained model is shown.

Realization of the human skill in the peg-in-hole task using hybrid architecture

Since impedance control can achieve the desired dynamics between tool and environment, it is widely used for complex tasks. But deciding the impedance parameters for the task is very difficult because they cannot be calculated analytically. If we can extract impedance parameters from human demonstration, it is appropriate to use them. However, if there exist disturbances and/or noise at the playback stage, which was not taken into account at the skill acquisition stage, the executed task by the robot is far different from the human demonstration. In order to realize the robust skill, we introduce a hybrid architecture. This architecture enables robots to use the most suitable dynamics (impedance parameters) for each task situation. Moreover, because of event driven architecture, we can expect the improvement of robustness of the skill in the time domain. We apply our proposed method to the peg-in-hole task, and show some experimental results.

Implementation of the human skill based on hybrid architecture

Impedance control has been widely used in various manufacturing processes. In the impedance control framework, it is a very important and difficult problem of how to decide impedance parameters to realize the prescribed desired task. Since there are large similarities between impedance control and human fingertips control, one promising technique to overcome this problem is extracting the impedance parameters from human demonstration, and implementing them on the robot controller. However, if there exist disturbances and/or noise at the playback stage, which are not taken into account at the skill acquisition stage, the executed task by robot is far different from the human demonstration. In order to realize the robust skill, the authors introduce the hybrid architecture. This architecture enables robots to use the most suitable dynamics (impedance parameters) for each task situation. Moreover, because of event driven architecture, one can expect the improvement of robustness of the skill in the time domain.

Design of hybrid controller for assembly task based on evaluation with HMM

This paper presents a new design methodology for a hybrid controller in skill control problem. The proposed method is based on the performance evaluation with HMM. Our hybrid controller switches an impedance parameter, which is designed by mean: of referring the human demonstration, according to the task state (contact configuration). An event observer, which detects a change of state, plays an essential role in our hybrid controller. The performance of the event observer is evaluated by comparing the difference between the working data generated by human worker and the robot controlled by the hybrid controller involving the designed event observer. The difference is measured based on HMM distance. Some experimental results as for a connector insertion task are shown to verify the usefulness of the proposed method.