Motion Strategy Generation Based on Multimodal Motion Primitives and Reinforcement Learning Imitation for Quadruped...

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Task-oriented reinforcement learning (RL) has advanced quadruped robot capabilities.
• Current control strategies demand significant expertise and time for skill acquisition and multi-task control.
• Managing complex behaviors with coordinated actions remains a challenge.

Purpose of the Study:

• To propose a novel motion policy generation method for quadruped robots.
• To address limitations in current control strategies for complex behaviors and multi-task motion control.
• To simplify the design process and accelerate the deployment of operational skills.

Main Methods:

• Developed a multimodal motion library using 3D engine design, motion capture retargeting, and trajectory planning.
• Designed a temporal domain-based behavior planner to combine motion primitives for complex behaviors.
• Implemented a reinforcement learning (RL)-based imitation learning framework for precise trajectory tracking and rapid policy deployment.

Main Results:

• The proposed method effectively generates complex behaviors by combining multimodal motion primitives.
• The RL-based imitation learning framework enabled precise trajectory tracking and rapid policy deployment.
• Simulations and physical experiments on the Lite3 quadruped robot validated the approach's efficacy.

Conclusions:

• The study presents a new paradigm for developing and deploying motion strategies for quadruped robots.
• The method simplifies the acquisition of operational skills and enhances multi-task motion control.
• This approach offers a more efficient and effective way to manage complex robotic behaviors.