Exploration-based learning of a stabilizing controller predicts locomotor adaptation.

Area of Science:

• Biomechanics
• Robotics
• Neuroscience

Background:

• Human locomotion adapts seamlessly to body and environmental changes.
• Mechanisms for improved performance (e.g., energy efficiency, symmetry) and fall avoidance during adaptation are not fully understood.

Purpose of the Study:

• To model human locomotor adaptation as an interaction between a fast stabilizing controller and a slow reinforcement learning process.
• To predict and explain adaptation phenomena across various conditions like split-belt walking and exoskeleton use.

Main Methods:

• Developed a computational model integrating a reactive stabilizing controller with a reinforcement learner.
• The reinforcement learner uses local exploration and memory to optimize performance.
• Model predictions were validated against ten prior experiments and two new model-guided experiments.

Main Results:

• The model accurately predicts time-varying adaptation in diverse scenarios, including split-belt treadmills, asymmetric leg weights, and exoskeleton use.
• It captures key learning and generalization phenomena observed in human locomotion.
• Energy minimization with a small asymmetry cost emerged as a key performance measure.

Conclusions:

• A model combining reactive control and reinforcement learning provides a unified framework for understanding locomotor adaptation.
• This approach can explain performance improvements and stability maintenance during adaptation.
• The findings offer insights for designing better rehabilitation strategies and controlling wearable robots.