An Interpretable Neural Control Network With Adaptable Online Learning for Sample Efficient Robot Locomotion Learning

Area of Science:

• Robotics
• Machine Learning
• Artificial Intelligence

Background:

• Reinforcement learning for robot locomotion faces challenges with sample inefficiency and lack of interpretability.
• Existing methods often operate as 'black boxes', hindering analysis and improvement.

Purpose of the Study:

• To develop a novel, sample-efficient, and interpretable framework for robot locomotion learning.
• To address the limitations of traditional reinforcement learning in robotic applications.

Main Methods:

• Introduction of the Sequential Motion Executor (SME), an interpretable three-layer neural network for motion generation.
• Implementation of the Adaptable Gradient-weighting Online Learning (AGOL) algorithm to prioritize relevant parameter updates.
• Integration of SME and AGOL to create an analyzable learning framework.

Main Results:

• SME-AGOL achieved 40% sample reduction compared to state-of-the-art methods.
• Demonstrated a 150% increase in final reward and locomotion performance on a simulated hexapod robot.
• Achieved efficient learning on a physical hexapod robot within 10 minutes from scratch.

Conclusions:

• The proposed SME-AGOL framework offers a sample-efficient and understandable approach to robot locomotion learning.
• Interpretability in learning frameworks can be leveraged to enhance both sample efficiency and overall performance.
• This work paves the way for more transparent and effective reinforcement learning in robotics.