MIT’s clumsy cheetah robot taught itself to sprint with the help of AI

Mini Cheetah, a four-legged robot from the Massachusetts Institute of Technology, learned to navigate all types of terrain during a strictly virtual training protocol.

Robotics is one of those fields of research that has made impressive progress in recent years. In particular, we can mention Boston Dynamics, an American company that distinguished itself with such incredible machines as the humanoid acrobat Atlas or Spot, the dog that is today one of the international stars of robotics. But today we are talking about another cyber quadruped; here Mini Cheetah (“cyber cheetah”)a quadruped from MIT who just set a pretty impressive personal best.

The first thing that strikes you when you see the Mini Cheetah in development is that it looks surprisingly clumsy. Indeed, the least we can say is that it falls far short of the grace and finesse that Spot is capable of, let alone its biological counterparts; instead, he squirms wildly like a puppy promised a treat.

This is not a programming error or a lack of skill on the part of the MIT teams, far from it; if his movements seem so strange, it’s because the various movements were never programmed separately. Indeed, the special feature of Mini Cheetah is that he learned to run on his own!

Completely self-taught

Indeed, before gaining the right to a physical presence, Mini Cheetah spent the first moments of his life exclusively in virtual form. 4,000 digital versions of the ship encouraged exploring their anatomy to learn how to navigate its environment.

And they had to do it from scratch, without a single example of movement from a real animal; this is called reinforcement learning. It’s a bit like a child being encouraged to reach for a rattle; The goal is to encourage him to explore all the physical possibilities his body gives him so he can find ways to use it to move.

Therefore, each model developed its own way of walking and then running; Then the MIT researchers were able to synthesize the most promising findings into a final model that allowed them to test the fruits of their labor on a real physical body.

Another fundamental difference from many robots: Mini Cheetah does not have any camera that would allow it to observe its environment. He completely relies on his feelings and reflexes developed during training; relying on his memory, he can adapt his course to more or less flat, slippery terrain, etc.

The Spot is generally more impressive, but it’s completely incapable of sprinting like this. ©Boston Dynamics

Shabby looking, but damn effective

Admittedly, the end result doesn’t exactly exude elegance; but in practice it is extremely strong, stable and versatile. Therefore, it is able to move in environments where other four-legged robots of the same type would have difficulty putting one foot in front of the other.

Despite its name, the Mini Cheetah is also not as fast as its biological counterparts. If the latter can easily exceed a hundred kilometers per hour; even if it just broke its own record, the robot is still happy to approach 15 km/h. A significant number for a robot of this type; but that performance remains well below that of another cheetah robot made by Boston Dynamics, whose top speeds are comparable to those of Usain Bolt.

But the most interesting thing is, above all, the combination of these two last points in one work; it is both unmanned and fast, which is a real paradigm shift. Nothing to do with the slow and careful approach taken by most modern prototypes. And this is largely thanks to self-learning with AI; this approach made it possible to optimize the training time, ignoring everything that did not allow the work to move faster.

And too bad if it looks clumsy as long as the result is there! “Instead of a human determining exactly how the robot should walk, it learns from its own experience in a simulator to learn how to move very, very fast.”, explains Gabriel Margolis, one of the researchers associated with the study.

Free yourself from the conceptual limitations of people

What we see here is one of the main interests of machine learning based systems; they are very effective in solving the specific tasks set before them,” explains Tonnes Nygaard, a robotics researcher interviewed by Wired. “In this case, the algorithm finds the fastest way to make the robot work, even if it looks shaky“, he clarifies.

But above all, this approach is not only functional, but also faster than development »old“. Indeed, it would simply be impossible for engineers to individually program all the scenarios the Mini Cheetah encounters during virtual self-learning.

And this approach can be applied to almost any system affected by machine learning; if we are willing to sacrifice all practical and ergonomic constraints, AI is more likely to achieve its goal. And behind this observation there may even be a small life lesson: sometimes it is extremely important to think outside the box.

Human Seekers are limited by their own idea of ​​what constitutes a “good” race. These criteria may be based on old traditions, on the work of other researchers, on nature, or even on a subconscious preference for symmetry or beauty.Nygaard told Wired. “But these criteria often limit our approach and produce worse results!“, he concludes. On good terms!

Leave a Comment

Your email address will not be published. Required fields are marked *