Real-time prediction learning for the simultaneous actuation of multiple prosthetic joints

Area of Science:

• Robotics
• Biomedical Engineering
• Artificial Intelligence

Background:

• Prosthetic control systems can be enhanced by integrating learned predictions.
• Real-time learning and prediction maintenance are crucial for advanced prosthetic functionality.

Purpose of the Study:

• To investigate the benefits of incorporating temporally extended, real-time predictions into prosthetic control systems.
• To demonstrate the first combination of actor-critic reinforcement learning with real-time prediction learning for prosthetic control.

Main Methods:

• Developed and evaluated a novel control learning approach combining actor-critic reinforcement learning with real-time prediction learning.
• Tested the system during the myoelectric operation of a robot limb.

Main Results:

• The integrated approach potentially speeds up control policy acquisition and enables unsupervised adaptation in myoelectric controllers.
• Demonstrated facilitation of synergies in highly actuated limbs and enabled anticipatory actuation for coordinated motion.
• Provided initial evidence for the practicality of real-time prediction learning in complex prosthetic systems.

Conclusions:

• Real-time prediction learning is a practical method to support intuitive joint control in advanced prosthetic systems.
• This approach can lead to faster learning, unsupervised adaptation, and improved coordination in prosthetic devices.