Autonomously learning to visually detect where manipulation will succeed

摘要: Visual features can help predict if a manipulation behavior will succeed at given location. For example, the success of that flips light switches depends on location switch. We present methods enable mobile manipulator to autonomously learn function takes an RGB image and registered 3D point cloud as input returns which is likely succeed. With our methods, robots train pair support vector machine (SVM) classifiers by trying behaviors locations in world observing results. Our require change state between two sets (e.g., switch up down), detect when each has been successful, initial hint where one be successful. When feature associated with location, trained SVM predicts successful To evaluate approach, we performed experiments PR2 robot from Willow Garage simulated home using flip switch, push rocker-type operate drawer. By active learning, efficiently learned SVMs enabled it consistently these tasks. After training, also continued order adapt event failure.