Substitute Tool Select Method using Contact Force Data Trained Neural Network Model

Taro Takahashi, Yutaro Ishida, Naoki Mima, Keisuke Takeshita, Yoshiaki Asahara, Hiroshi Bitc, Hideki Kajima

Abstract: In this study, we propose a method for selecting substitute tools using force data. The majority of investigations in this area have employed geometric data. However, it is possible to select tools with the same function but different shapes if the substitute tools can be estimated using force data. For example, if a robot intends to hit something with a hammer, it can find a substitute tool, such as a stone instead of a hammer. Our approach first measures the force between the original tool and manipulated objects via tele-operated simulation and then evaluates candidate substitute objects based on their ability to match the demonstrated force profile. The force data at each contact point on a three-dimensional tool was projected onto a two-dimensional force map, and then the force map was input to a classification head using a multilayer perceptron to find a substitute tool as a classification problem. Considering the high dimensionality of the force map, we also investigated the application of a force map encoder using an auto-encoder or convolutional neural network. These three types of substitute tool estimation were performed using a striking motion with nine candidate objects, resulting in a recall value of over 78%.