Can robots run?
In recent years, we have seen the creation of increasingly complex robotic animals with incredible abilities. Taking inspiration from the animal world, scientists have managed to create sophisticated and innovative machines.
From robotic devices capable of perform flips up to those able to straighten up after tripping (and passing through the big "war dogs"), we have really seen everything.
Yes, but can they run?
Yet there is one animal skill that still remains particularly difficult to reproduce: running.
Robots, in fact, are unable to run by automatically adjusting their pace when moving from one environment to another, and are unable to maintain a sustained speed without falling.
MIT's new Mini Cheetah could be the exception.
All 'Improbable AI Lab at MIT, a team of researchers attempted to train a robot on all fours to run, using a machine learning approach. In fact, he attempted to teach him how to learn to run.
Did he succeed? Judge for yourself. Thanks to training, the robot reached speeds of almost 15 kilometers per hour (8,7 miles per hour).
The new “Cheetah Robot”
Mini Cheetah's locomotion is regulated through the use of certain algorithms, developed by a group of engineers.
The team's work has been and continues to be fundamental, as well as very complex. Algorithm development can take longer than expected, even for the most trivial fixes. Months of hard work were needed to teach the robot the best way to “run”.
Relying on a machine learning model, the Mini Cheetah is able to update itself autonomously every time it faces a formative experience. As if he were equipped with a "human memory", he remembers his mistakes and plans not to make them in the future.
Fortunately, the scientists got to prepare it properly. The quadrupedal robot can in fact learn through experiences on real and simulated terrain.
Thanks to this merger, the robot was able to acquire 100 days of real-world experience in hours.
A revolutionary method
The learning method chosen by MIT scientists transforms the robot into a sort of "little child". Just as happens with newborns, as mentioned, the Mini Cheetha learns to run from its mistakes and builds a safe path in which to move.
The simulated environment makes everything safer, avoiding real risks and excessive dangers.
At the moment, Mini Cheetah is the only quadruped bot to have learned to run almost autonomously. Unfortunately, however, it is not the fastest in the world. Just think of the robot Out Runner six-legged dog that can run at 70 kilometers per hour (45 mph) on a treadmill, or several others.
What really matters in this project is not the speed, but the learning technique.
MIT researchers plan to open source soon this algorithmo, so as to allow the start of other experiments.
Who will be able to make these "so" run even more?