On Automated Object Grasping for Intelligent Prosthetic Hands Using Machine Learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Biomedical Engineering

Background:

• Current prosthetic technology faces challenges in autonomous grasping and user experience.
• Extensive user training is often required for fine motor control in electronic prosthetics, limiting usability.
• Improving prosthetic autonomy is crucial for enhanced functionality and acceptance.

Purpose of the Study:

• To propose an automated method for prosthetic grasping control.
• To leverage computer vision and machine learning for enhanced prosthetic autonomy.
• To evaluate the performance of different reinforcement learning algorithms in automated grasping tasks.

Main Methods:

• Utilized computer vision-based techniques and machine learning algorithms.
• Employed three reinforcement learning algorithms: Soft Actor-Critic (SAC), Deep Q-Network (DQN), and Proximal Policy Optimization (PPO).
• Trained agents for automated grasping tasks using these algorithms.

Main Results:

• The Soft Actor-Critic (SAC) algorithm demonstrated the highest success rate at 99% within 200,000 timesteps.
• Object physical characteristics were found to influence the agent's ability to learn an optimal grasping policy.
• SAC outperformed DQN and PPO in achieving successful autonomous grasping.

Conclusions:

• The Soft Actor-Critic (SAC) algorithm shows significant potential for developing intelligent prosthetic hands.
• This automated method can lead to prosthetic hands with automatic object-gripping capabilities.
• The findings suggest a path towards more intuitive and autonomous prosthetic control.