Self-organizing neural network for reproducing human postural mode alternation through deep reinforcement learning

Area of Science:

• Robotics and Control Systems
• Computational Neuroscience
• Human Motor Control

Background:

• Postural coordination involves switching between in-phase and anti-phase modes, a phenomenon crucial for stability during activities.
• Existing models often rely on prior body knowledge, limiting adaptability to changing body characteristics or unknown dynamics.
• Learning mechanisms are vital for developing adaptive neural networks for robust postural control.

Purpose of the Study:

• To create a self-organizing neural network capable of adaptive postural mode coordination without a predefined body model.
• To investigate the network's ability to reproduce and transition between in-phase and anti-phase coordination modes.
• To assess the network's hyper-adaptivity to dynamic task conditions and altered body mass.

Main Methods:

• A deep reinforcement learning algorithm was employed to train a neural network for postural coordination.
• The study utilized head-target tracking tasks with varying target frequencies to induce transitions between coordination modes.
• Performance was evaluated using correlation and relative phase analysis of hip and ankle joints.

Main Results:

• The self-organizing neural network successfully reproduced in-phase and anti-phase postural coordination modes.
• Transitions between these modes were achieved by altering head-target tracking task frequencies.
• The trained network demonstrated adaptability to continuous task variations and unlearned body mass changes, maintaining mode alternation.

Conclusions:

• Self-organizing neural networks can learn adaptive postural control strategies, mimicking human motor adaptability.
• The developed model offers a pathway for creating more robust and versatile robotic systems and understanding biological motor control.
• This approach highlights the potential of deep reinforcement learning in achieving hyper-adaptive motor control for dynamic environments.