ANYmal parkour: Learning agile navigation for quadrupedal robots.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Agile navigation in four-legged robots is complex due to dynamic motions, robot-environment contact, and limited sensor views.
• Existing methods often require expert demonstrations, offline computation, or prior environmental knowledge.

Purpose of the Study:

• To develop a fully learned approach for agile navigation in four-legged robots capable of handling complex, parkour-like scenarios.
• To enable robots to autonomously select and control locomotion skills for diverse obstacles.

Main Methods:

• A hierarchical reinforcement learning framework was employed, training specialized locomotion skills (walking, jumping, climbing, crouching).
• A high-level policy was developed to select and adapt skills based on the environment and robot capabilities.
• A perception module was trained to reconstruct obstacles from noisy, occluded sensor data for scene understanding.

Main Results:

• The proposed method successfully navigated challenging, parkour-like scenarios without expert demonstrations or prior environmental knowledge.
• The system demonstrated effective skill selection and adaptation, enabling robust obstacle traversal.
• Real-world experiments showed successful transfer from simulation to hardware, with robots navigating obstacles at speeds up to 2 meters per second.

Conclusions:

• The fully learned, hierarchical approach significantly advances agile navigation capabilities in four-legged robots.
• The method offers a robust and adaptable solution for complex robotic navigation tasks, overcoming limitations of previous approaches.
• Successful real-world deployment validates the approach's effectiveness and potential for autonomous robotic systems.