Learning Robotic Grasp using Visual-Tactile model

摘要

The practice of grasping an object in humans, depend greatly on the feedback from tactile sensors. Nevertheless, the recent works of grasping in robotics, has been constructed only from visual input, but in this case the feedback after instigating contact cannot be easily benefited. A survey is done and presented in this paper to explore how the tactile information is used by the robot to learn to adjust its grasp proficiently. Additionally, an action-conditional model which uses raw visual- tactile data that learns grasping strategies is presented. The model presented iteratively selects the most favorable actions which implements the grasp. The approach does not require any analytical modeling of contact forces nor calibration of the tactile sensors, thereby decreasing the engineering requirements for obtaining a competent grasp strategy. The model, a two-finger gripper with tactile sensors of high-resolution on each finger was trained with data from various grasping trials. After a number of rigorous testing, it was seen that the approach had effectively learned useful and interpretable grasping behaviors. To conclude, the selections made by the model were studied and it was seen that it had effectively learned suitable and apt behaviors for grasping.