MIT's Small Cheetah automaton learns to run on its own


You may also like this partner content (after ad)

The old adage that you learn best from your mistakes just happened to hold true for quadruped robots. The Small Cheetah, developed by MIT, saw itself implement the self-learning system that allowed it to learn to run on its own which .

For the time, his movements still seem a little clumsy. Sober in fact, the MIT video demonstrating the progress of Small Cheetah, or Small Gupard, is even so unappealing as it reminds us of an overexcited puppy discovering the world. However, his abilities have something to impress.

In top place, because running, for a metal man, is not obvious. To obtain a fast training course, it is necessary to push the equipment to its limits, for example by operating near the maximum output torque of the motors. In such circumstances, the dynamics of the automatic robot is difficult to model analytically. Robotics must react quickly to changes in the environment, such as the minute it encounters ice while running on the grass

, analyze scientists at MITs Dubious AI Laboratory , Gabriel Margolis and Ge Yang.

Small Cheetah nevertheless managed to record a sober speed report at 3.9 meters per second (11 km/they would), according to the MIT. It is also able to rotate at high speed. The 2nd interesting stage in this efficiency is that the software learned to run by itself. Sober effect, the scientists used the neural network as a learning master.

This machine learning algorithm is called neural network because it is originally inspired by the biological neurons. The idea is to feed the algorithm with a lot of data so that it can study results by comparison and thus learn according to what works or not. Small Cheetah therefore learned in the simulator, and in just three hours, she gained experience in a fantastic number of possible situations, which she could then reuse in the real world. )

In short, the neural network has done the work sober those who usually program robots: develop how an automaton should act in all the possible circumstances is simply very difficult. The process is tedious, as if a metal man were to fail on a particular surface, a human engineer would have to identify the result of the failure and manually adapt the robot’s controller, and this process may require considerable human effort. Trial-and-error learning eliminates the human need to specify precisely what robotics should behave in every circumstance

, explain Gabriel Margolis and Ge yang.

Reduce the robot learning temperature

Scientists therefore claim that this approach could greatly simplify the development of sober robots: The traditional paradigm in robotics is that humans tell software once what to do and how to do it. The problem is that such a framework is not scalable, as it would take an immense amount of human engineering work to manually code an automaton with the necessary skills to operate in many diverse environments. A more practical way to design el metal man with many various skills is to terrible at automatic robot what to do and let him understand opinion

. Bet rather successful, since we can see on the video that Small Cheetah indeed adapts to mishaps terrain, steps, gravel or even snow.

The MIT video:

Supply: MIT Information