Learning Quadrupedal High-Speed Running on Uneven Terrain

Area of Science:

• Robotics
• Control Systems
• Artificial Intelligence

Background:

• Reinforcement learning (RL) controllers are used for quadruped robot locomotion on uneven terrain.
• High-speed movement amplifies external disturbances, impacting robot stability.
• Existing RL methods require high control frequencies, leading to significant computational costs.

Purpose of the Study:

• To propose a novel control framework that reduces computational cost while maintaining stability for high-speed quadruped robot locomotion.
• To address the latency challenge in RL-based control systems for dynamic robotic applications.

Main Methods:

• A hybrid control framework combining a low-frequency RL-based control policy with a high-frequency model-based joint controller.
• The RL policy outputs joint control laws, executed by high-frequency controllers to mitigate disturbances.
• Evaluation conducted on simulated terrains with varying height differences (up to 6 cm).

Main Results:

• Achieved a running speed of 1.8 m/s with the Unitree A1 quadruped in simulation.
• The RL policy operates at 50 Hz with 20 ms latency, while the joint controller runs at 1000 Hz.
• Demonstrated effective mitigation of external disturbances and latency issues.

Conclusions:

• The proposed framework successfully overcomes latency limitations of low-frequency RL updates.
• This approach is applicable for real-world deployment of high-speed quadruped robots.
• Enables stable and efficient locomotion on challenging terrains.