Model-free reinforcement learning for robust locomotion using demonstrations from trajectory optimization.

Area of Science:

• Robotics
• Machine Learning
• Control Theory

Background:

• Deploying policies on real robots often requires extensive retraining.
• Existing methods struggle with environmental uncertainties and dynamic tasks.

Purpose of the Study:

• To develop a general, two-stage reinforcement learning approach for creating robust robot policies.
• To enable policy deployment on real robots with minimal or no additional training.

Main Methods:

• A two-stage reinforcement learning framework utilizing a single demonstration from trajectory optimization.
• Stage one: Demonstration used for initial exploration.
• Stage two: Direct task reward optimization for robust policy computation.

Main Results:

• The approach successfully generated robust policies for dynamic hopping and bounding tasks on a quadruped robot.
• Demonstrated effective policy deployment on real robots without further training.
• Validated robustness against environmental uncertainties.

Conclusions:

• The proposed two-stage reinforcement learning method is effective for creating generalizable and robust robot policies.
• This approach significantly reduces the need for real-world robot retraining.
• It shows promise for dynamic locomotion tasks in robotics.