This robot dog can learn to walk alone in just an hour

Often, these are instead robots inspired by the living; but here, this relationship goes in two ways, which is very interesting for the researchers.

In wild animals, infant animals can benefit from rapid growth, especially those far above the food chain. For example, if a small giraffe can’t learn to run fast, it has every chance to be a snack to the first predator to arrive; natural selection therefore pushes them to learn to walk in just a few hours.

Granted, their actions are far from perfect at this stage; they are clumsy, poorly coordinated, but this adaptation allows them to increase their chances of passing the decisive stage in early childhood. Excluding human children, who remain extremely vulnerable for many months.

For researchers in robotics and artificial intelligence, these are very interesting examples; Understanding the ins and outs of these mechanisms could allow them to build more efficient engines and very fast driving systems.

So researchers at the renowned Max Planck Institute in Germany are investigating it. They created a quadruped robot that seemed obscure at the Boston Dynamics Spot in an effort to figure out how animals learned to walk.

We seek answers by creating a robot that displays reflexes, like an animal, and learns from its mistakes. “, explained Felix Ruppert, one of the researchers associated with this work.

A robotic version of the body pacemaker

What sets this robot apart is that the researchers equipped it with a system that, frankly, plays the role of “ virtual spinal cord “. In humans, this structure that occupies the canal of the spinal column has a network of more specific neurons: the spinal locomotor networks (or CPG, for central pattern generators).

It is a one-of-a-kind tangle of nerve cells that can function completely autonomously. In summary, very little conductor in the nervous system ; it is he who sets the stage for a whole set of behaviors based on the so -called “ rhythmic and stereotyped “.

Examples of patterns seen in electrical signals generated by CPGs. © Marder and Bucher

They are able to act independently of any other nerve impulses; for example, if you can breathe, blink, or walk carelessly, it’s thanks to CPGs.

But the situation will change when simple automation is no longer enough, for example when the ground becomes uneven. Here, the model offered by CPGs is no longer sufficient. The rest of the musculoskeletal system comes into play to get the job done thanks to various reflex functions.

Also in the case of walking, it is thanks to these reflexes that you don’t have to end up with your nose on the floor at every wrong step; the nervous system will automatically play to contract the muscles to compensate for movement and maintain posture.

In animals, including humans, learning to walk is like constantly “training” these CPGs and reflexes; Researchers describe them as a ” automatic and integrated gait intelligence “. And it is for imitating this extremely flexible neurological architecture that the researchers chose to integrate a virtual CPG into their robot.

A robot that will have fun in just an hour

A set of sensors located on its legs allows this system to be trained permanently. On each return, the algorithm compares the actual position of the limbs in the model proposed to the real-time CPG, then calculates the additional movements needed to compensate for the difference.

The ” manual »Suggested by the CPG purifies over time, which makes the robot more comfortable when walking. ” Initially, when the computer generates a signal that controls the leg motors, the robot stumbles. “, the researchers will explain.

Data travels between the virtual GC and leg sensors and is continuously compared. If the sensor data does not match expectations, the algorithm will change some parameters until the robot stops tripping, etc. “, they determined.

And this method works so well that Morti can learn to walk one about an hour, without any prior information! A shocking time passed, knowing that it was traditionally counted in days or months for other robots of this class.

This graph represents the efforts that the robot must give to avoid a fall. They quickly fall over the turns, showing that the robot has learned to walk very fast. © Rupert et al.

If robots serve as models to the living

It’s already a huge advantage in itself, but it’s not the only one. Evolutionarily speaking, one of the great things about GICs is that they allow us to perform relatively complex actions without having to spend a lot of energy actively thinking about them. And we also see this advantage of this cybernetic approach.

Where such robots always eat ten or even hundreds of wattsMorti makes a very reasonable energy budget in 5W. Knowing that autonomy is still one of the main limitations of these machines, this is a remarkable improvement.

The final advantage of this method is that it can even benefit biologists! In fact, as a general rule, roboticists are inspired by living animals to optimize their machine. But here, it can also be reversed. This robot can provide a very interesting research platform for researchers working on the subtleties of the nervous system.

You can’t easily study the spinal cord of a living animal, but you can model one on a robot. “, explained Alexander Badri-Spröwitz, head of the laboratory and co-author of the study.” We know that these GPCs are present in many animals. We also know what reflexes are mixed, but we don’t fully understand the interactions between the two. “, he determined.

This is fundamental research at the intersection of robotics and biology “, he was enthusiastic about the conclusion.” This robotic model offers us answers that biology alone cannot provide “. The circle is complete!

The text of the study is here.

Leave a Comment