The Mini-Cheetah cheetah robot, which has long been developed by specialists from the Biomimetic Robotics Lab and Improbable AI Lab at the Massachusetts Institute of Technology (MIT), recently set its own speed record of 14.04 kilometres per hour, thanks to a new control and self-learning system based on artificial intelligence principles. The presence of such a system allows the robot to choose its own way of moving, adapting quickly enough to changes in the nature of the surface and requiring no human intervention or involvement.
But even the robot equipped with artificial intelligence, Mini-Cheetah, cannot yet be called the fastest. In 2012, its larger counterpart from the famous Boston Dynamics company set a speed record for this type of robot, which was 45.54 kilometres per hour. The previous achievement, however, was the result of careful work and calculations by humans, which took into account absolutely all parameters and features of the robot's construction. The Mini-Cheetah robot, however, has the potential to become even faster and to do this completely independently by learning, experience and skill.
In the video below, you can see how the Mini-Cheetah robot behaves when encountering obstacles, when walking with one defective limb, when navigating on slippery ice surfaces and over chaotically formed piles of gravel. All this is made possible by a fairly simple artificial neural network that independently evaluates the current situation, the results obtained earlier and selects, based on this, the most optimal movement.
Moreover, you do not even need to create real polygons with such conditions to teach the robot Mini-Cheetah the wisdom of movement in different environments. The robot can then be trained in a virtual reality environment that can simulate even the most exotic conditions. Moreover, such learning takes place much faster than in the real world; in three hours of intensive simulation, a robot is capable of gaining experience equivalent to that gained in the real world over 100 days of continuous training.
The technology developed by the Massachusetts scientists could also be successfully applied to other types of robots, which will be able to acquire highly versatile capabilities that are only achieved with great difficulty using the traditional approach.
"The most practical way to create a robot with a wide variety of skills and abilities is to allow the robot to be given an end result and let it figure out for itself how to achieve what it wants," the researchers write, "In our work, we have used this paradigm in relation to other automated systems, including manipulators that can grasp, move and perform other actions on various objects."