They are born with muscular coordination networks with rigid threads located in their spinal cord that they depend on to stand and move through neural reflexes.
Although a little more basic, motor control reflexes help a puppy avoid falling and hurting themselves during their first attempts at walking. Then, further and more precise muscle control is practiced, until the nervous system is well adapted to the muscles and tendons of the animal’s legs so that it can keep pace with adults.
Researchers at the Max Planck Institute for Intelligent Systems (MPI-IS) in Stuttgart, Germany, conducted a study (published in the journal Intelligence of natural machines) to find out how animals learn to walk and overcome obstacles. For this, they built a four-legged robot, the size of a medium-sized dog, which helped them understand the details.
“As engineers and robotics professionals, we seek the answer by building a robot that thinks like an animal and learns from its mistakes,” said Felix Ruppert, a former doctoral student in the Dynamic Locomotion Research Group at MPI- IS, in a statement. “If an animal stumbles, is it a mistake? Not if it happens once. But if it stumbles often, that gives us a measure of how well the robot walks.
A learning algorithm optimizes the virtual spinal cord of a robot dog
A Bayesian optimization algorithm drives machine learning: information measured by the foot sensor is combined with target data from the modeled virtual spinal cord running as a program on the robot’s computer. It learns to walk by continuously comparing information sent and received from the sensor, running loops reflexes and adapting their motor control patterns.
The learning algorithm adapts the control parameters of a Central Pattern Generator (CPG). In humans and animals, these central pattern generators are neural networks in the spinal cord that produce periodic muscle contractions without brain input.
Networks of central pattern generators help generate rhythmic tasks such as walking, blinking, or digesting. Additionally, reflexes are involuntary motor control actions triggered by encoded neural pathways that connect sensors in the leg to the spinal cord.
As the young animal walks on a perfectly flat surface, CPGs may be sufficient to monitor motion signals from the spinal cord. A small bump on the ground, however, changes the course. Reflexes kick in and adjust movement patterns to prevent the animal from falling.
These momentary changes in motion signals are reversible, or “elastic,” and motion patterns return to their original configuration following the disturbance. But if the animal does not stop stumbling through many cycles of movement – despite active reflexes – then the movement patterns must be repaired and made “plastic”, i.e. irreversible.
In the newborn animal, the CPGs are initially not yet snug enough and the animal stumbles, both on flat and uneven ground. The animal, however, quickly learns how its CPGs and reflexes control the muscles and tendons of the legs.
The same goes for the robot dog called “Morti”, which, in addition, optimizes its movement patterns faster than an animal, in regarding an hour. Your CPG is simulated on a small, lightweight computer that controls the movement of your legs.
The virtual spinal cord is placed on the quadruped robot where the head would be. During the hour it takes the robot to walk smoothly, data from its foot sensor is continuously compared to the expected stumble predicted by the CPG.
If this happens, the learning algorithm changes the distance and speed at which the legs swing back and forth and measures how long it stays on the ground. Adjusted movement also affects how the robot can use its compatible leg mechanics. During the learning process, the CPG sends suitable motorized signals so that the robot starts to stumble less and optimizes its walking.
In this structure, the virtual spinal cord has no explicit knowledge regarding the design of the robot’s legs, its motors and its springs. Knowing nothing regarding the physics of the machine, he lacks a “model” robot.
“Our robot was practically ‘born’ knowing nothing regarding its autonomous legs or how they work,” says Ruppert. “The CPG is like an integrated automatic walking intelligence that nature provides and that we transfer to the robot. The computer produces signals that control the leg motors, and the robot initially walks and stumbles. Data travels up from the sensors to the virtual spinal cord, where the sensor and CPG data are compared. If the sensor data does not match the expected data, the learning algorithm modifies the walking behavior until the robot walks well without tripping. Changing the CPG output, keeping reflexes active, and monitoring robot stumbles, is central to the learning process.
Low power consumption makes the mechanism more viable
While four-legged industrial robots from well-known manufacturers, which have learned to operate using complex controllers, are very power-hungry, Morti’s computer only needs five watts to run.
“We cannot easily search the spinal cord of a living animal. But we can model one on the robot,” says Alexander Badri-Spröwitz, co-author of the study and head of the Dynamic Locomotion Research Group, in Max Planck. “We know that these CPGs exist in many animals. And we know that reflexes are built in, but how can we combine the two so that animals learn movements with reflexes and CPGs? This is fundamental research at the intersection of robotics and biology. The robotic model gives us answers to questions that biology alone cannot answer.
Have you watched our new videos on Youtube? Subscribe to our channel!
