The focus of this research is to develop a method for teaching quadrupedal robots to move effectively on deformable terrain.

      Advancements in robotics and machine learning have led to the development of legged robots that can navigate complex environments, including uneven and soft terrains. However, while simulation-based reinforcement learning approaches have proven to be effective in controlling legged robots, their performance on soft and deformable terrains has been limited, particularly at high speeds. This is due to the fact that reinforcement learning models are only effective within the data distribution they have been trained on, and therefore, cannot perform well in environments that they have not previously encountered

      To address this challenge, a team of researchers has introduced a versatile and computationally efficient granular media model for reinforcement learning that can be parameterized to represent diverse types of terrain, ranging from very soft beach sand to hard asphalt. Additionally, the team has developed an adaptive control architecture that can implicitly identify the terrain properties as the robot feels the terrain, and then use this information to improve the robot's locomotion performance. 

        The team applied their techniques to the Raibo robot, a dynamic quadrupedal robot developed in-house. Through training the robot's neural networks, they were able to achieve high-speed locomotion on deformable terrains, even when the robot's feet were completely buried in the sand during the stance phase. The robot achieved a speed of 3.03 meters per second on soft beach sand, demonstrating its ability to run at high speeds on difficult terrain.

*

*

*

*