Learning Grasp Configuration Through Object-Specific Hand Primitives for Posture Planning of Anthropomorphic Hands

Area of Science:

• Robotics
• Biomechanics
• Machine Learning

Background:

• Controlling anthropomorphic hands with multiple degrees of freedom presents challenges.
• Postural synergy theory offers a new framework for hand control.
• Generating grasp configurations for novel tasks remains a significant hurdle.

Purpose of the Study:

• To develop a method for learning grasp configurations based on object shape using postural synergy theory.
• To enable robotic hands to adapt grasping strategies for new tasks.

Main Methods:

• An experimental paradigm was designed to record human hand joint angles during grasping and operational tasks with 50 objects.
• Principal Component Analysis (PCA) was used to extract four hand primitives and establish a low-dimensional synergy subspace.
• Gaussian Mixture Regression (GMR) and Gaussian Processes (GPs) were employed to learn and infer synergy inputs for trajectory planning.

Main Results:

• A low-dimensional synergy subspace was successfully established, transforming joint trajectory planning into synergy input determination.
• The proposed method demonstrated the ability to generate grasp configurations for prosthetic hand control.
• Simulations showed that the method can achieve human-hand-like grasping movements, extending from simple to complex tasks.

Conclusions:

• The developed method effectively learns grasp configurations from object shapes using postural synergy theory.
• This approach enhances the adaptability and dexterity of robotic hands, particularly for prosthetic applications.
• The findings suggest a promising direction for advanced human-robot interaction and control systems.