They stumble around unsteadily: newborn animals are clumsy at first. A robot dog now simulates how they then learn to use their legs skillfully: “Morti” can optimize his movement patterns using an adaptive computer system that acts as an artificial nervous system. In just one hour, the robot learns to move fluently on its own. According to the scientists, the concept can thus be used for basic research at the interface between robotics and biology.
It becomes particularly clear with a foal: Newborn animals first have to learn to use their muscles and tendons in a coordinated manner in order to be able to move forward confidently. In the early attempts to walk, initially only stabilizing movements based on reflexes protect the young animals from violent falls. The finer muscle control, on the other hand, has to be developed first. Basically, the following emerges: The so-called central movement pattern generator in the spinal cord is trained during the bumpy attempts at walking until the movement control can finally enable the young animal to have good control over the legs. But how exactly this happens is unclear.
“We cannot study the spinal cord of a living animal very well. But it can be modeled in the robot,” says Alexander Badri-Spröwitz from the Max Planck Institute for Intelligent Systems in Stuttgart. “We basically know that there is a central movement pattern generator in animals and reflexes. But how are both combined in such a way that they can learn with the reflexes and the movement pattern generator?” In order to develop a system for researching this question, Badri-Spröwitz and his colleague Felix Ruppert built the four-legged robot “Morti”. “It’s a system that has reflexes like an animal and learns from mistakes,” says Ruppert.
Learning algorithm trains a virtual spinal cord
Morti is equipped with a movement pattern generator in the form of a small computer, which controls the artificial muscles and tendons in his legs and thus corresponds to his biological model. In humans and animals, networks of nerve cells in the spinal cord make up this system, which causes rhythmic muscle contractions without the influence of the brain. As long as a young animal is on the move without any problems, this controller sends its basal movement signals unchanged. This changes when the animal stumbles. Then reflexes kick in and adjust the movement pattern so it doesn’t fall. If the animal repeatedly stumbles through processes that are fundamentally suboptimal, learning processes can lead to a permanent adaptation of the movement patterns, the researchers explain.
They were now able to give Morti exactly this ability as well. This was made possible by a so-called Bayesian optimization learning algorithm, which influences the development of the movement pattern generator, which was initially not optimally adjusted. Sensor information from the feet is compared with the target data from the computer – the virtual spinal cord. The robot then learns to walk better and better by continuously adapting the structure of the movement patterns sent to the sensor information, the scientists explain. In concrete terms, this means that if the robot stumbles, the learning algorithm changes how far the legs swing back and forth, how fast they move and how long a leg stays on the ground. The “updated” movement pattern generator then sends signals that enable the robot to walk with as little stumbling as possible: Morti gradually learns to walk more and more skilfully.
Confident walk after an hour
As it turned out, the robot is even slightly superior to its biological role models: in contrast to young animals, Morti develops a safe locomotion pattern from its initially bumpy walk on the treadmill within an hour. The optimization is also reflected in the energy four-belly when running, the scientists report: By making better use of the advantages of its mechanics, the robot improves its energy efficiency by 42 percent. Ruppert and Badri-Spröwitz have thus succeeded in developing a model of how animals learn to walk. “Our robot is born symbolic and knows nothing about how its legs work. Our system works like a built-in automatic walking intelligence, which nature provides us with and which we have transferred to the robot,” Ruppert sums up.
As the two scientists explain, the system can now be further developed and thus be used for basic research at the interface between robotics and biology. Because their system is obviously interesting from a technical point of view as well: the computer in Morti consumes only five watts. The researchers say that other adaptive systems with comparable performance require significantly more energy. Above all, however, Morti depicts nature: “The robot model can give us answers to questions that biology alone cannot answer,” says Badri-Spröwitz.
Source: Max Planck Institute for Intelligent Systems, specialist article: Nature Machine Intelligence, doi: 10.1038/s42256-022-00505-4