Enhancing Dexterous Hand Control: a Distributed Architecture for Machine Learning Integration

Purpose The aim of this paper is to enhance the control performance of dexterous hands, enabling them to handle the high data flow from multiple sensors and to meet the deployment requirements of deep learning methods on dexterous hands. Design/methodology/approach A distributed control architecture was designed, comprising embedded motion control subsystems and a host control subsystem built on ROS. The design of embedded controller state machines and clock synchronization algorithms ensured the stable operation of the entire distributed control system. Findings Experiments demonstrate that the entire system can operate stably at 1KHz. Additionally, the host can accomplish learning-based estimates of contact position and force. Originality/value This distributed architecture provides foundational support for the large-scale application of machine learning algorithms on dexterous hands. Dexterity hands utilizing this architecture can be easily integrated with robotic arms.