Learning robust autonomous navigation and locomotion for wheeled-legged robots.

Area of Science:

• Robotics
• Artificial Intelligence
• Urban Systems

Background:

• Urban autonomous navigation faces challenges in varied terrains and dynamic obstacles.
• Wheeled-legged robots offer potential for enhanced logistics and adaptability.
• Existing systems often lack integrated locomotion and navigation control.

Purpose of the Study:

• To develop and validate a fully integrated system for autonomous wheeled-legged robot navigation in urban environments.
• To enhance robot adaptability and efficiency in complex, real-world settings.
• To demonstrate the effectiveness of hierarchical reinforcement learning for robot control.

Main Methods:

• Utilized model-free reinforcement learning (RL) and privileged learning for a versatile locomotion controller.
• Implemented a hierarchical RL framework integrating locomotion and navigation controllers.
• Developed mobility-aware local navigation and large-scale path planning modules.
• Validated the system through autonomous, kilometer-scale navigation missions in urban areas.

Main Results:

• Achieved efficient and robust locomotion over diverse rough terrains with seamless mode transitions (walking/driving).
• Demonstrated effective high-speed navigation through challenging terrain and dynamic obstacles.
• Successfully completed autonomous navigation missions in Zurich and Seville, validating system robustness and adaptability.

Conclusions:

• Integrated control systems are crucial for seamless autonomous navigation in complex urban environments.
• Wheeled-legged robots and hierarchical RL are feasible for advanced autonomous navigation.
• The developed system has significant implications for last-mile delivery and future logistics.