Deep learning-based control framework for dynamic contact processes in humanoid grasping

Abstract

Humanoid grasping is a critical ability for anthropomorphic hand, and plays a significant role in the development of humanoid robots. In this article, we present a deep learning-based control framework for humanoid grasping, incorporating the dynamic contact process among the anthropomorphic hand, the object, and the environment. This method efficiently eliminates the constraints imposed by inaccessible grasping points on both the contact surface of the object and the table surface. To mimic human-like grasping movements, an underactuated anthropomorphic hand is utilized, which is designed based on human hand data. The utilization of hand gestures, rather than controlling each motor separately, has significantly decreased the control dimensionality. Additionally, a deep learning framework is used to select gestures and grasp actions. Our methodology, proven both in simulation and on real robot, exceeds the performance of static analysis-based methods, as measured by the standard grasp metric Q1. It expands the range of objects the system can handle, effectively grasping thin items such as cards on tables, a task beyond the capabilities of previous methodologies.