A novel approach to locomotion learning: Actor-Critic architecture using central pattern generators and dynamic motor...

Area of Science:

• Robotics
• Computational Neuroscience
• Bio-inspired AI

Background:

• Robot locomotion learning is challenging due to diverse morphologies and gait requirements.
• Existing methods often struggle with adaptability and generalization across different robot types.
• Bio-inspired approaches offer promising solutions for robust and adaptive locomotion.

Purpose of the Study:

• To propose and evaluate a novel bio-inspired controller architecture for learning diverse locomotion gaits.
• To investigate the effectiveness of combining central pattern generators (CPGs) and dynamic motor primitives (DMPs) for locomotion.
• To compare different actor-critic algorithms for optimizing locomotion adaptation in robots.

Main Methods:

• Developed a three-layer bio-inspired controller architecture integrating CPGs for baseline motion and DMPs for dynamics adaptation.
• Employed actor-critic algorithms (policy search and policy gradient) to learn the "reshaping" function for gait adaptation.
• Tested the architecture on a humanoid robot learning to crawl and a puppy robot learning to gallop.

Main Results:

• The proposed architecture successfully enabled learning of distinct locomotion gaits (crawling, galloping) for different robot morphologies.
• Actor-critic algorithms demonstrated effective learning of the dynamics adaptation "reshaping" function.
• Comparison of policy search and policy gradient revealed varying advantages for different robot morphologies.

Conclusions:

• The bio-inspired controller architecture provides a versatile framework for robot locomotion learning.
• The combination of CPGs and DMPs offers a powerful approach for adaptive and generalized locomotion.
• The study contributes a generic architecture and insights into algorithm selection for robot gait learning.