A Stability Training Method of Legged Robots Based on Training Platforms and Reinforcement Learning with Its...

Area of Science:

• Robotics
• Control Systems
• Machine Learning

Background:

• Legged robots require robust stability control for dynamic environments.
• Existing controllers often struggle with unpredictable disturbances and varied robot morphologies.
• Developing adaptable self-stabilization is crucial for autonomous robotic operation.

Purpose of the Study:

• To introduce a learning-based controller, the global self-stabilizer, for achieving self-stabilization in legged robots.
• To validate the controller's effectiveness on a human-sized biped robot (GoRoBoT-II) under various perturbations.
• To demonstrate the controller's ability to dynamically adapt actions based on system state.

Main Methods:

• The global self-stabilizer was designed with three modules: action selection, adjustment calculation, and joint motion mapping.
• Learning algorithms were developed for each module of the controller.
• Simulations and experiments were conducted on the GoRoBoT-II using a motion platform with tilt and impact perturbations.

Main Results:

• The global self-stabilizer demonstrated convergence after training.
• The controller successfully combined actions dynamically based on the robot's system state.
• Stability verification success rates increased by over 20% in simulations and 15% in experiments compared to baseline controllers.

Conclusions:

• The proposed global self-stabilizer is a feasible and effective approach for enhancing legged robot stability.
• The learning-based controller offers significant improvements in robustness against external disturbances.
• This method holds potential for application to legged robots of various sizes and leg configurations.