Hierarchical Vision Navigation System for Quadruped Robots with Foothold Adaptation Learning.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Legged robots offer dynamic locomotion capabilities for navigating complex terrains.
• Efficient navigation and dynamic utilization in cluttered environments remain significant challenges for legged robots.
• Current systems often struggle with real-time adaptation to unpredictable obstacles and terrain variations.

Purpose of the Study:

• To develop a novel hierarchical vision navigation system for quadruped robots.
• To enhance navigation efficiency and dynamic control in cluttered and dynamic environments.
• To enable robots to autonomously adapt foothold placement for robust locomotion.

Main Methods:

• A hierarchical system combining a high-level end-to-end navigation policy with a low-level foothold adaptation policy.
• The high-level policy generates optimal paths with obstacle avoidance.
• The low-level policy utilizes auto-annotated supervised learning for adaptive locomotion control and feasible foot placement.

Main Results:

• The system demonstrates efficient navigation in dynamic and cluttered environments.
• Successful obstacle avoidance and target approach were achieved without prior environmental information.
• Experiments in both simulation and real-world scenarios validated the system's effectiveness.

Conclusions:

• The proposed hierarchical vision navigation system significantly improves navigation efficiency for quadruped robots.
• The integration of adaptive foothold planning enhances robot's ability to handle challenging terrains.
• This approach offers a robust solution for autonomous navigation in unstructured and dynamic environments.