A robot dog learns to walk on its own in an hour

⇧ [VIDÉO] You might also like this partner content (after ad)

How do animals’ reflexes allow them to learn to walk in such a short time? Researchers from the Max Planck Institute for Intelligent Systems have looked into the matter. To answer this, however, they used robotics rather than biology. For the study, they designed a robot that could learn to walk, without prior instruction, in just one hour.

This is how “Morti” was born, a dog-robot designed to study the development of walking in young animals. “ Researching the spinal cord of a living animal is complex. But using a robot, we can model one “Explains Alexander Badri-Spröwitz, who co-wrote the publication, in a press release from the Institute. The research was published in the journal nature machine intelligence.

Of course, scientists already know a lot about the motor functioning of animals. Indeed, as the statement recalls, animals are born with muscle coordination networks embedded in their spinal cord. They therefore learn very quickly to walk to escape predators. At the beginning, they rely on these somewhat basic data, but which allow them to start moving. Subsequently, they learn to coordinate muscles and tendons more precisely. Their movements gradually become more fluid, but it takes practice.

Scientists claim that these two steps are driven by two mechanisms. In humans and animals, there are what they call “ central pattern generators (CPG), i.e. networks of “neurons” in the spinal cord that produce periodic muscle contractions without the intervention of the brain. This mechanism occurs for very repetitive tasks, such as blinking, walking, etc. These are basic tasks, which do not require to be really connected to the brain. In addition to this, the “reflexes” intervene. These are involuntary movement actions, the scientists explain, which are also hardwired directly to the spinal cord, not the brain.

A virtual spinal cord

Concretely, on a flat surface without pitfalls, the central pattern generator works. Reflexes come into play when there is something unusual, a dip, a bump, etc. In the newborn animal, the CPGs are initially not yet snug enough and the animal stumbles, whether on flat or uneven ground. Over time, he learns how his reflexes and “core patterns” control his movements, and improves. In other words, experience in the field gradually corrects pre-existing data on walking. However, knowing the details of this operation is not easy. The scientists therefore decided to go through robotics.

a: Photograph of Morti. b: 2D rendering of Morti on a treadmill. Morti was constrained to the sagittal plane by a linear rail and lever-guide mechanism that allowed pitching of the body around its center of gravity. Inset: Close-up cross-section of the FootTile sensor on Morti’s foot segment. The polyurethane sensor dome deforms under load and the pressure sensor measures the increasing pressure in the air cavity to detect foot contact. © Felix Ruppert et al.

As engineers and roboticists, we searched for the answer by building a robot with reflexes like an animal and learning from its mistakes. », explains Felix Ruppert, one of the authors. So they paired a basic program with a machine learning algorithm. In other words, they started by creating a virtual “spinal cord”, which they attached to the head of the robot dog. The researchers didn’t provide Morti with any basic information about his body shape, motors, springs…

The learning algorithm therefore made it possible to compare the sensory information perceived with the basic program of the robot to gradually improve its capacities. The robot dog designed by the researchers proved capable of learning, in the same way as animals, to walk: and even faster than an animal, since it only took him an hour to get to good coordination of their movements. During this time, the sensor data from the feet (paws) was continuously compared to the landing predicted by the robot’s CPG. When the robot stumbled, the algorithm adapted the swing speed of the legs, the distance, the duration of contact with the ground, etc.

We know that these CPGs exist in many animals. We know that reflexes are integrated; but how do you combine the two so that animals learn movements with reflexes and CPGs? This is fundamental research at the intersection between robotics and biology. The robotic model allows us to answer questions that biology alone cannot answer “say the scientists.

Source: Nature Machine Intelligence

Leave a Comment