Whatever the nature of the ground, this quadruped robot adapts its gait on its own so as not to fall and this is a real feat. It can even run at 11 km/h on sand.
The gymnastic prowess and abilities of the robot Atlas from Boston Dynamics impress and his videos systematically make the buzz. The robot can jump, perform somersaults and other stunts, with flexibility, but it is very far from certain that Atlas can walk in the sand without falling. Evolving on uneven ground, and in particular sandy, remains a feat for a robot. It is on this subject that a team of researchers from the department of mechanical engineering of the Korean Higher Institute of Science and Technology Kaist in South Korea has worked.
The scientists began by modeling the stresses exerted by walking in the sablesable, before experimenting with it with a quadruped robot called RaiBo. What is natural for us requires collecting a lot of data and having controllers that will allow us to immediately adjust the approach to the constraints detected. Thus, for the robot, when its paw lands on a sandy surface, the Artificial Intelligence (AI) must predict the depth and the vitessevitesse into which it sinks in order to be able to immediately compensate for the movementmovement. To achieve this, the modelizationmodelization initial had to be very close to the real ground. This equivalent of our “proprioception” also allowed him to stand and turn on himself on a mattress. pneumaticpneumatic.
The robot filmed pivoting on itself on an air mattress. The quadruped had to rely on its network of artificial neurons to interpret the nature of the surface and adapt the pace by making decisions in real time. © Kaist
A race at 11 km/h in the sand
In practice, it was machine learning that then made it possible to strengthen the robot’s analysis and reaction capacities on the different terrain surfaces that it practised. In the end, it could evolve stably by adapting to the characteristics of the ground without the need to revise its control algorithm. In other words, the AI reinforced its own learning by training itself to improvise in order to be able to walk in a stable manner, whatever the nature of the ground. Even better, the robot was able to move at a speed of 11 km/h on a sandy beach. The results of these experiments were published in the magazine Science Robotics.
