It takes a bit of imagination to recognize a dog in the four-legged machine designed by Max Planck researchers in Stuttgart. However, the analogy to a living being is important to the scientists at the Institute for Intelligent Systems, so they speak of a robot dog the size of a Labrador, which they even gave a name to: “Morti”.
“Morti” is not only about the development of advanced robot technology, but also about the biological question of how young animals learn fluid movement sequences. A dog – or a foal or a giraffe – is born with fully equipped motor skills and can stand and walk on four legs almost immediately.
But everyone knows the videos of newborn baby animals, in which you can see them stumbling and stumbling around awkwardly at first. Apparently, the precise control of the muscles and the movement sequences still have to be learned through real experience. It is clear that the movement of the legs is controlled by the nervous system in the spinal cord.
The Max Planck researchers have equipped “Morti” with a “virtual spinal cord” and are now reporting in the journal “Nature Machine Intelligence” how the robot independently learned to walk smoothly on its legs. Animals have a genetically determined basic equipment that enables them to take their first steps without falling and possibly injuring themselves. Over time, they learn and improve their movements.
The researchers also equipped “Morti” with basic “reflexes” and an artificial intelligence that enables the machine to learn from mistakes. It is an adaptive algorithm that has the function of the nervous system in the spinal cord and is not yet perfectly adjusted at the beginning.
Unlike flesh and blood beings, electronics can adapt much faster than biological neural networks. The researchers report that after just one hour, the robotic dog has optimized the coordination of its legs.
“We cannot study the spinal cord of a living animal,” says study co-author Alexander Badri-Spröwitz. “But we can model it in the robot.” In this way, the scientists pursued the research question of how animals learn to walk from stumbling blocks.
“As engineers or roboticists, we sought the answer by building a robot that has reflexes like an animal and learns from mistakes,” says lead author Felix Ruppert. “If an animal stumbles, is that a mistake? One time not. But if he stumbles a lot, that gives us a measure of how well the walking is going.”
The researchers installed sensors on the feet of “Morti” with which the actual data on the movement sequence are obtained. The target data for good movement are stored in the virtual spinal cord, which by the way is located in the robot where the head would be in a dog.
The robot dog learns to walk by continuously changing the structure of the control signals to the movement mechanics in such a way that the actual data adapts to the specified target data.
For example, if the robotic dog stumbles, the algorithm changes certain parameters – how far the legs swing back and forth, how fast they swing, or how long a leg stays on the ground. If this leads to an improvement, the selected parameters are adopted.
“This is basic research at the interface between robotics and biology,” says Badri-Spröwitz. “The robot model gives us answers to questions that biology alone cannot answer.”
