Advanced biomimetic robotic hand with EMG lifelong learning and recognition

Area of Science:

• Robotics
• Humanoid Robot Design
• Biomedical Engineering

Background:

• Designing robotic hands for humanoid robots, especially toddler-sized ones, presents significant challenges.
• Controlling robotic hands effectively using biological signals requires sophisticated optimization and learning techniques.

Purpose of the Study:

• To optimize the design of an anthropomorphic robotic hand for a toddler-sized humanoid robot.
• To develop a control system for the robotic hand using surface electromyographic (sEMG) signals.

Main Methods:

• Employing Isolation Forest Backward Particle Swarm Optimization to refine the robotic hand's design, focusing on thumb opposability and grasping ability.
• Utilizing Learning Without Forgetting (LWF) for sequential training of sEMG data to create an ensemble model for robotic hand control.
• Simulating grasping scenarios in Webots to validate and optimize the robotic hand design.

Main Results:

• The optimized robotic hand demonstrated superior performance, achieving the highest fitness values in both simulation and real-world tests compared to existing designs.
• Comparison of sEMG data processing techniques (raw, bandpass, discrete wavelet transformed bandpass) within the LWF framework identified optimal input for neural network structures.
• Successful real-time manipulation of the robotic hand via hand gesture classification using the final LWF model in a practical system.

Conclusions:

• The study successfully optimized an anthropomorphic robotic hand design and developed an effective sEMG-based control system.
• The integration of advanced optimization algorithms and sequential learning techniques provides a robust solution for robotic hand control in humanoid robots.
• The developed system shows promise for real-world applications requiring intuitive and precise robotic hand manipulation through biological signals.