Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : June 08, 2016 - June 11, 2016
In most of the past works, a robotic picking is tackled in simulated or structured real world. In a clutter real environment, however, it is still difficult to complete for robots, because of uncertain vision input of environment and characteristic of objects themselves. In this paper, we propose a learning approach while the robot is doing the task to be adaptive to the uncertainty of vision data from environment which the designer of the robotic system had not estimated. We approach to this issue by estimating perception parameter from the vision input at each time, and updating that estimator while the task. Firstly, the robot collects data from environement and initialize the estimator for initial value of parameters which the designer assigned. Secondly, it does the task and collects data of whether the task is completed in what environement status. And finally, the estimator is updated after training with the collected data. We adopted the approach to two perception components of picking task in clutter scene: perception to segment objects in clutter scene, and that to recognize objects in hand after picked while moving arms. The recognition accuracy rose in both segmentation and in-hand recognition. Our approach is a general way to estimate better parameter for the task in an environement status, and efficient to be adaptive the change of environment.
