A scientific team from the Max Planck Institute for Intelligent Systems (MPI-IS) has equipped the robot dog Morti with an algorithm that enables it to learn to walk from mistakes and successes within an hour. Sensors provide the algorithm with the information it needs to optimize a central movement pattern generator.
The scientists at the MPI-IS based their research on how newborn animals learn to walk. They can’t use their legs properly until they can coordinate the muscles and tendons perfectly. Reflexes help to survive the first attempts at walking unscathed. By practicing the movements, the nervous system in the spinal cord is continuously adapted until precise muscle control is possible, the scientists describe the starting point of their research work in the paper “Learning plastic matching of robot dynamics in closed-loop central pattern generators” published in Nature Machine Intelligence .
The video shows how the robot dog Morti gradually learns to walk.
The team built a robotic dog the size of a Labrador. In order to give the four-legged friend the necessary reflexes so that the robot can learn from its mistakes, they installed sensors on its feet. The movement data from the sensors is evaluated by an algorithm and compared with the signals from a Central Pattern Generator (CPG).
Optimized movement patterns
In a way, the CPG corresponds to what the nervous system in the spinal cord of animals does. For example, if Morti stumbles on uneven ground, the CPG receives new information from the algorithm about how to move the legs and how to position the legs in order not to fall. From then on, the CPG sends the adapted signals for improved leg movement. The system continuously optimizes the movement patterns.
The CPG, which is simulated by a small computer, has no information whatsoever about the structure of the robot and its mechanics. This is similar to an animal that is also born unaware of its body but still learns to walk.
“The CPG works like a built-in automatic walking intelligence that nature provides us with and which we transferred to the robot. The computer produces signals that go to the legs. The robot runs and stumbles as it does so. This data flows back from the sensors above and are compared by the computer, which generates the movement patterns blindly. If the data doesn’t meet the specifications, then we change the walking behavior until walking works without stumbling. Changing the specification of what the legs should do is the learning process,” says Felix Ruppert, former doctoral student of the research group Dynamic Locomotion at the MPI-IS.
However, robot dog Morti optimizes its movement patterns much faster than a newborn animal. It only takes about an hour, as the researchers describe. Compared to other robot systems, which can consume up to several hundred watts of power for the elaborate and complex control, the Mortis system works frugally with just 5 W.
To home page