Policy gradient optimization of controllers for natural dynamic mono-pedal gait

Area of Science:

• Robotics
• Control Systems
• Biomechanical Engineering

Background:

• A biologically-inspired dynamic controller for biped locomotion was previously proposed.
• This controller uses torque pulses timed to an internal phase variable.
• Controller parameters can be fixed (open-loop) or adapted (closed-loop).

Purpose of the Study:

• To implement and optimize this controller for a mono-ped robot.
• To enhance the robot's robustness using policy gradient optimization.
• To investigate policy transfer from simulation to hardware.

Main Methods:

• Implementation on a mono-ped robot.
• Policy gradient applied in simulation for parameter optimization.
• Learning a grounded action transformation (GAT) for sim-to-real transfer.

Main Results:

• Policy gradient successfully optimized controller parameters.
• GAT improved the correlation between simulation and experimental results.
• The optimized controller enhanced the mono-ped robot's performance and robustness.

Conclusions:

• The biologically-inspired controller is effectively implemented and optimized for a mono-ped robot.
• Policy gradient and GAT are valuable tools for robotic controller development and sim-to-real transfer.
• This approach significantly improves robot performance and robustness.