Learning quadrupedal locomotion on deformable terrain.

Area of Science:

• Robotics
• Machine Learning
• Control Systems

Background:

• Simulation-based reinforcement learning (RL) advances legged robot control but struggles with soft, deformable terrains due to distribution shift.
• Existing RL agents perform poorly in environments outside their training data distribution, limiting real-world applicability.
• High-speed locomotion on challenging terrains remains a significant hurdle for legged robots.

Purpose of the Study:

• To develop a versatile and computationally efficient granular media model for RL.
• To create an adaptive control architecture capable of identifying terrain properties during locomotion.
• To enhance the high-speed locomotion capabilities of legged robots on diverse and deformable terrains.

Main Methods:

• Introduced a parameterized granular media model adaptable to various terrain types (e.g., sand, asphalt).
• Developed an adaptive control architecture that implicitly identifies terrain properties via robot-environment interaction.
• Integrated terrain parameter identification to optimize locomotion control policies.
• Applied the techniques to a dynamic quadrupedal robot (Raibo) for training and testing.

Main Results:

• Achieved high-speed locomotion on soft beach sand at 3.03 m/s, even with feet fully buried.
• Demonstrated successful generalization to various terrains including vinyl tile, athletic track, grass, and an air mattress.
• The adaptive control architecture effectively utilized identified terrain parameters to improve performance.

Conclusions:

• The proposed granular media model and adaptive control architecture significantly improve RL-based legged robot locomotion on deformable terrains.
• This approach overcomes the data distribution limitations of traditional RL, enabling robust performance across diverse environments.
• The findings pave the way for more versatile and capable legged robots in complex, real-world scenarios.