Learning hybrid locomotion skills-Learn to exploit residual actions and modulate model-based gait control

Area of Science:

• Robotics
• Machine Learning
• Control Theory

Background:

• Legged robots require advanced control for stability against external perturbations.
• Existing methods often struggle with unexpected dynamic disturbances.

Purpose of the Study:

• To develop a hybrid framework combining machine learning and control for enhanced robotic balance.
• To improve stability and robustness of legged robots under external perturbations.

Main Methods:

• Developed a hybrid framework with a model-based gait kernel and a neural network.
• The neural network adjusts gait parameters and generates compensatory joint actions.
• Optimized seven Neural Network policies using residual actions for modulation and compensation.

Main Results:

• The hybrid framework significantly improved stability against external forces by up to 118%.
• Demonstrated robustness against measurement noise and model inaccuracies.
• Validated generalization to dynamic walking in unseen scenarios.

Conclusions:

• The proposed framework effectively enhances legged robot stability and robustness.
• Combining gait parameter modulation with residual actions is key to improved performance.
• The system shows strong potential for real-world applications requiring dynamic balance.