ARS: AI-Driven Recovery Controller for Quadruped Robot Using Single-Network Model.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Legged robots, particularly quadruped robots, excel in navigating challenging terrains.
• Falling is a common occurrence for legged robots, necessitating robust recovery mechanisms.
• Existing methods for fall recovery in robots are often complex and limited in scope.

Purpose of the Study:

• To develop a method for full recovery of quadruped robots from a fallen state.
• To utilize a single neural network model for autonomous fall recovery.
• To demonstrate the effectiveness of the proposed method on a physical robot platform.

Main Methods:

• A single neural network model was designed for fall recovery.
• The model was trained in two stages using reinforcement learning in simulation.
• The trained model was directly deployed on the AiDIN-VIII quadruped robot (12 degrees of freedom).

Main Results:

• The robot successfully recovered from various fall postures within 5 seconds.
• Complete recovery was achieved even when the robot was fully overturned.
• The method demonstrated effectiveness in recovering from externally induced disturbances.

Conclusions:

• A single neural network model can enable quadruped robots to achieve full recovery from falls.
• The proposed reinforcement learning approach is effective for training robust fall recovery policies.
• The method enhances the operational reliability and autonomy of legged robots in unstructured environments.