人間型ロボットの触覚制御に関する研究

点击翻译按钮获取中文摘要

As we have pointed out in "Work package 3 : Tactile object recognition" in the application for the JSPS postdoctoral fellowship, while interesting work on object recognition through tactile sensing has been presented in the past, there are limitations : usually the objects are fixated, so that they do not move during exploration, and they are oriented in a certain direction ; often grippers, and not dexterous hands are used ; if distributed pressure data is used, it comes from flat tactile arrays. The hands of TWENDY-ONE provide richer tactile data than any other robotic hand. It is equipped not only with distributed tactile skin sensors on most of the hand and 6-axis F/T sensors in each fingertip, but includes also somatic sensors such as motor angles and spring displacements. We have used that multifingered hand and normal grasping actions for tactile object recognition. The objects are allowed (and indeed expected) to move between grasping actions. When using tactile sensors, it is not clear what kinds of features are useful for object recognition. Recently, deep learning has shown promising results. Nevertheless, deep learning has rarely been used in robotics and to our best knowledge never for tactile sensing, probably because it is difficult to gather many samples with tactile sensors. We have employed deep learning techniques for tactile object recognition. The robot had to identify 20 different objects, the most challenging set ever used for tactile object recognition. Our results show a clear improvement when using a denoising autoencoder compared to traditional neural networks. We achieved a recognition rate of about 88%. This is one of the highest recognition rates reported so far for recognizing grasped objects with unknown orientation and translation relative to the hand. The results have also been submitted to the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2014 conference.