Fine-Tuning Myoelectric Control Through Reinforcement Learning in a Game Environment

Area of Science:

• Biomedical Engineering
• Neuroscience
• Robotics

Background:

• Myoelectric controllers for bionic prosthetics face challenges in accurately decoding motor intent.
• Current Supervised Learning (SL) methods require high-quality labeled muscle activity data, which is difficult to obtain during real-world use.
• Improving the reliability of these controllers is crucial for advanced prosthetic limb functionality.

Purpose of the Study:

• To investigate the potential of Reinforcement Learning (RL) to enhance motor intent decoding in myoelectric controllers.
• To incorporate usage-based electromyographic (EMG) data for improved controller performance.
• To overcome limitations of traditional SL methods in acquiring representative training data.

Main Methods:

• A pre-trained SL control policy using static EMG data was fine-tuned with RL.
• Dynamic EMG data was collected during interaction within a custom-designed game environment.
• Real-time experiments were conducted to evaluate the RL-enhanced approach.

Main Results:

• The RL-based method demonstrated effective prediction of simultaneous finger movements.
• Decoding accuracy during gameplay saw a two-fold increase.
• A 39% improvement in accuracy was observed in a separate motion test.

Conclusions:

• Reinforcement Learning (RL) significantly improves the accuracy and robustness of myoelectric controllers.
• Incorporating usage-based EMG data during RL fine-tuning is key to enhanced performance.
• This approach shows great promise for advancing the reliability of bionic limbs.