MIT's clumsy robot cheetah learned to sprint on its own using AI

MIT’s clumsy robot cheetah learned to sprint on its own using AI

Mini Cheetah, the quadruped robot from MIT, learned to move on all types of terrain during a strictly virtual training protocol.

Robotics is one of those fields of research that have made spectacular progress in recent years. We can notably cite Boston Dynamics, an American company which has distinguished itself with incredible machines such as the humanoid acrobat Atlas, or Spot, the dog which is today one of the international stars of robotics. But today, it is another cyberquadruped in question; here is Mini Cheetah (“cyber cheetah”)a quadruped from MIT who has just set a rather impressive personal record.

The first thing that strikes you when you see Mini Cheetah evolve is that it seems surprisingly clumsy. Indeed, the least we can say is that it is far from possessing the grace and finesse of which Spot is capable, not to mention its biological counterparts; instead, he wriggles frantically like a puppy being 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 is because the different movements have never been programmed individually. Indeed, the particularity of Mini Cheetah is that he learned to run on his own!

A fully self-taught robot

Indeed, before having the right to a physical presence, Mini Cheetah spent the first moments of its life in exclusively virtual form. 4,000 digital versions of the craft were encouraged to explore their anatomy to learn how to move in its environment.

And they had to do it from scratch, without any example of movement from a real animal; this is called reinforcement learning. It’s a bit like a baby being encouraged to trudge by shaking a rattle; the objective is to encourage him to explore all the physical possibilities offered to him by his body so that he finds himself how to use it to move around.

Each model has therefore developed its own way of walking, then of running; MIT researchers were then able to synthesize the most promising findings into a final model that was allowed to test the fruit of its labor with a real physical body.

Another fundamental difference compared to many robots: Mini Cheetah does not have any camera that would allow it to observe its environment. He relies entirely on his feelings and on the reflexes he has developed during training; by drawing on his memory, he can adapt his course to more or less regular, slippery terrain, and so on.

Wobbly in appearance, but damn effective

Admittedly, the end result doesn’t exactly exude elegance; but in practice it is extremely strong, stable and versatile. It is therefore able to move in environments where other quadruped robots of the same type would have a hard time putting one leg in front of the other.

Despite its name, 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 has just broken its own record, the robot is still content to approach 15 km/h. An impressive figure for a robot of this type; but this performance remains significantly lower than that of another cheetah robot produced 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 last two points in a single robot; it is both self-driving and fast, which is a real paradigm shift. Nothing to do with the slow and careful approach adopted by most current prototypes. And that’s largely thanks to the self-learning enabled by AI; this approach allowed to optimize the training time by ignoring everything that did not allow the robot to move faster.

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

Emancipate oneself from the conceptual limits of humans

What we see here is one of the main interests of machine learning-based systems; they are very effective in solving the specific problems given to them”, explains Tønnes Nygaard, a robotics researcher interviewed by Wired. “In this case, the algorithm finds the fastest way to make the robot run, even if it looks wonky”, he specifies.

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 encountered by Mini Cheetah during its virtual self-training.

And this approach can be applied to almost any system affected by machine learning; if we agree to sacrifice all the practical and ergonomic constraints, the AI ​​will most likely end up achieving its goal. And there may even be a little life lesson behind this observation: sometimes it is absolutely essential to think outside the box.

Human seekers are limited by their own conception of what constitutes a “good” race. These criteria can be based on old traditions, on the work of other researchers, on nature, or even on a subconscious preference for symmetry or beauty.”, Nygaard tells Wired. “But these criteria often limit our approach and give poorer results!”, he concludes. On good terms!

Leave a Comment

Your email address will not be published.