Learning and Reusing Quadruped Robot Movement Skills from Biological Dogs for Higher-Level Tasks

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Model Predictive Control (MPC) is a classic but complex method for quadruped robot motion control, requiring precise dynamics models that limit agility.
• Model-free learning methods offer a promising alternative, reducing modeling difficulties and computational load for enhanced robot control.
• Biological systems provide inspiration for developing sophisticated motion control strategies.

Purpose of the Study:

• To propose a hierarchical reinforcement learning (HRL) framework for quadruped robot motion control.
• To enable robots to learn complex tasks by building upon basic motion skills, inspired by animal development.
• To achieve more agile, biomimetic movements and improve adaptability in challenging environments.

Main Methods:

• A hierarchical reinforcement learning framework is proposed, inspired by biological learning processes.
• Basic motion skills are learned from biological dog motion data.
• Domain randomization techniques are employed during training for robust policy transfer to physical robots.
• Higher-level tasks are learned using pre-learned basic skills, reducing redundant training.

Main Results:

• The proposed HRL framework facilitates the learning of higher-level motion tasks for quadruped robots.
• Trained policies demonstrate direct transferability to physical robots due to domain randomization.
• The controller enables more biomimetic movements, enhancing robot agility and adaptability.
• The approach allows for efficient operation in complex terrains by leveraging robot capabilities.

Conclusions:

• Hierarchical reinforcement learning offers an effective model-free approach for advanced quadruped robot motion control.
• The framework reduces development complexity and training time while improving performance.
• The method enhances biomimicry, agility, and adaptability, paving the way for robots in complex environments.