You’ve probably seen quadruped robots in movies, series or even TV news. They are becoming something more and more common in the robotics universe, and now researchers at MIT (Massachusetts Institute of Technology) have discovered how to radically improve the control of their movements. THE The answer lies in the use of artificial intelligence (AI) to teach them to walk and run.
This type of robot, which is still under development, could be useful in many situations. The four legs allow the machine to adapt to the most varied terrains and reach high speeds with stability. For example, these “robot dogs” could assist in rescue missions on Earth or even be used in the exploration of Mars. The possibilities are many.
Nonetheless, programming movements is time consuming and complex. Traditionally, software engineers have had to code countless responses and predict a huge variety of situations. They control robot leg movements with a set of rules.
For example, if the sensors detect a certain amount of force in leg y, the program will respond by starting motor a to perform another action to maintain stability, and so on.
That’s where artificial intelligence would come in as an alternative approach. AI uses machine learning to develop their own style of command execution, all based on trial and error tests.
The MIT researchers more specifically used a technique called “reinforcement learning”, which works by giving the AI an objective known as a “reward function”. In practice, a goal is set, such as “run as fast as you can”, and the program will try to figure out how to achieve this result from scratch on the basis of experimentation.
This means that the process of controlling robot dog movements will be faster using AI compared to manual programming. Furthermore, the machine becomes more adaptable to situations outside the engineers’ initial calculations.
Mini Cheetah sets new speed record
The MIT group created new software for the university’s quadrupedal research robot, called the Mini Cheetah. Using reinforcement learning, they managed to reach a new top speed for the machine, of 3.9 m / s, or about 14 km/h.
As you can see in the video above, the Mini Cheetah’s movements are still a little clumsy. However, this is the result of an AI that is only optimizing for speed, not thinking about running in a way that looks “natural” to human eyes.
“Reinforcement learning finds a way to run fast, but given an underspecified reward function, there is no reason to prefer a pace that is ‘natural’ or preferred by humans,” Gabriel Margolis, co-author of the research, told The Verge.
MIT engineers also pointed out that the great advantage of developing control software using AI is that it consumes less time than calculating all the specific physics of each terrain.
Using a simulator, researchers can place the robot in any virtual environment for the AI to test and calculate all variables without actually having to put it in the field. According to the MIT group, its artificial intelligence-enhanced simulator was able to accomplish in just three hours what would have taken a hundred days of tests in the real world.
This is another important step towards seeing increasingly intelligent robots coming into action. Soon, we will be able to see them roaming inhospitable environments such as volcanoes and caves, and even exploring Mars.
With information: The Verge, MIT