This robot taught itself to run fast

A team from MIT taught their Mini Cheetah to run. For this, it took a good dose of artificial intelligence and machine learning. In three hours, the robot has learned one hundred days of virtual routes to then adapt its speed of movement according to the ground and obstacles. Result, a speed record.

The speed movement is one of the big criticisms made of robots. Whether it’s a humanoid, Spot the dog or even recently Bex the ibexthe joints mechanized structures and its weight prevent the robot from moving quickly. Moreover, as the machine does not know the possible obstacles that it will cross, the robot is always in apprehension and analysis before taking a step. Another major obstacle: the cost of materials. Given the price of a robot, researchers avoid pushing it to its limits.

Except that the robotics researchers of the CSAIL of Massachusetts Institute of Technology (MIT) came up with the idea of ​​using simulations powered by Artificial Intelligence (AI) to quickly teach the robot to adapt its style walking as needed. In this case, it was their robot-panther to whom they wanted to learn to adapt its approach according to the terrain. Until now, they had taught him to run fast on a conveyor belt or perform somersaults. There, it is a question of teaching him to run in a natural environment.

Reinforcement learning

In just three hours, the robot, known as Mini Cheetah, has stored 100 days of virtual adventures on a wide variety of terrains. Only the “animal” has learned techniques to modify its gait according to the terrain and the textures always with the same objective: to get from point A to point B. As we can see in the video below, it’s bound to be fun.

So that the feline learns on its own, the researchers opted for reinforcement learning (reinforcement learning). This type of machine learning works with a system of rewards or penalties, or rather trial and error. The robot is unable to recognize the type of ground, whether dry or wet, smooth or irregular, but depending on the feeling of a step, it is able to adapt its stride and its movements. This is all the more effective when he makes several passes in the same place, and he can then move even faster.

A very good look

As a result, the panther reached a speed of 3.9 m/s, or 14 km/h. It’s faster than a normal human since, for a runner, it represents 4 min 15 sec per kilometer. You already have to be a very good runner to hold such a pace. This is a record for a robot autonomously, and the next step will now be to go further in behavioral autonomy, whether it is grasping objects or learning to move on unfamiliar ground. , without training with an AI.

” At the heart of AI research is the trade-off between what humans need to integrate (nature) and what the machine can learn on its own (build), explains an engineer from MIT. The traditional robotics paradigm is that humans tell the robot both what task to do and how to do it. The problem is that such a framework is not scalable, as it would take an immense human engineering effort to manually program a robot with the skills to operate in many diverse environments. A more practical way to build a robot with many diverse skills is to tell the robot what to do and let it figure out how. Our system is an example.

Our colleagues Gizmodo also mention the drone who, too, could take advantage of this type of learning to be “launched” into unknown spaces and yet fly autonomously.