Continual Learning for Anthropomorphic Hand Grasping

Abstract

It is important and challenging to learn to grasp different objects with anthropomorphic robotic hands continually and incrementally. However, most current works do not have this property: They learn grasp planners using large pre-prepared datasets, do not generalize well to new objects, and are difficult to improve continually. Besides, existing continual leaning works rarely target at anthropomorphic hand grasping, and usually deal with short streams of experiences. Because of the intrinsic long stream nature of anthropomorphic hand grasping, it is hard to utilize off-the-shelf continual learning methods for it. In this paper, we propose to introduce continual machine learning into anthropomorphic hand grasping and design the Continual Learning Framework of Anthropomorphic Grasping (CLFAG framework). It includes three modules: Data Producer, Grasp Experiences, and Continual Learning Algorithm ACL, thus makes the continual learning of anthropomorphic grasping possible. To overcome the catastrophic forgetting problem in long streams of grasping experiences, we propose a continual learning algorithm based on importance-based regularization and diversityaware replay within the CLFAG framework. Furthermore, we construct a dataset for continual learning of anthropomorphic grasping. Experiments on constructed dataset and in simulation demonstrate the effectiveness and superiority of the proposed approach.