A robotic dog with a virtual spinal cord can learn to walk

We’ve all seen those adorable clips of newborn giraffes or foals first learning to walk on their wobbly legs, stumbling around until they finally master the moves.

Researchers wanted to know how animals learn to walk and learn from their stumbles, so they built a four-legged, dog-sized robot to simulate it, according to a new study reported in Nature Machine Intelligence.

They found that it took their robot and its virtual spinal cord just an hour to gain control of its gait.

Getting up and walking quickly is essential in the animal kingdom to avoid predators, but learning leg muscle and tendon coordination takes time.

Initially, baby animals rely heavily on hardwired spinal cord reflexes to coordinate muscle and tendon control, while motor control reflexes help them avoid falling and injuring themselves on their first attempts.

Finer muscle control must be practiced until the nervous system adapts to the muscles and tendons and the young can then keep up with the adults.

“As engineers and roboticists, we sought the answer by building a robot that has reflexes like an animal and learns from mistakes,” says first author Dr. Felix Ruppert, former PhD student in the Dynamic Locomotion research group at the Max Planck Institute for Intelligent Systems (MPI-IS), Germany.

“If an animal stumbles, is that a mistake? Not once it happens. But if it stumbles a lot, that gives us a measure of how well the robot is running.”

Build a virtual spinal cord to learn to walk

The researchers designed a learning algorithm to act as the robot’s spinal cord and to work as a so-called Central Pattern Generator (CPG). In humans and animals, the CPGs are networks of neurons in the spinal cord that produce periodic muscle contractions without input from the brain.

These are important for rhythmic tasks like breathing, blinking, digestion and walking.

The CPG was simulated on a small and lightweight computer that controlled the movement of the robot’s legs, and it was positioned on the robot where the head would be on a dog.

The robot – which the researchers dubbed Morti – was equipped with sensors on its feet to measure information about its movement.

Morti learned to walk without first having explicit “knowledge” about his leg construction, motors, or springs by continuously comparing the expected data (modeled from the virtual spinal cord) with the sensor data while attempting to walk.

“Our robot is practically ‘born’ and doesn’t know anything about its leg anatomy or how it works,” explains Ruppert. “The CPG resembles a built-in automatic walking intelligence that nature provides, which we transferred to the robot. The computer generates signals that control the motors of the legs, and the robot walks and stumbles at first.

“The data flows back from the sensors to the virtual spinal cord, where sensor and CPG data are compared. If the sensor data does not match the expected data, the learning algorithm changes the walking behavior until the robot walks well and without stumbling.”

Sensor data from the robot’s feet is continuously compared to the expected touchdown data predicted by the robot’s CPG. If the robot stumbles, the learning algorithm changes how far the legs swing back and forth, how fast the legs swing, and how long a leg stays on the ground.

“Changing the CPG output while keeping the reflexes active and monitoring the robot’s stumbling is a key part of the learning process,” says Ruppert.

Morti the robot on the treadmill.
Morti the robot on the treadmill. Photo credit: Felix Ruppert, Dynamic Locomotion Group at MPI-IS

In an hour, Morti can go from stumbling around like a newborn animal to running, optimizing his movement patterns faster than an animal and increasing his energy efficiency by 40%.

“We cannot simply study the spinal cord of a living animal. But we can model one in the robot,” says co-author Dr. Alexander Badri-Spröwitz, Head of the Dynamic Locomotion Research Group.

“We know that these CPGs are found in many animals. We know that reflexes are embedded; but how can we combine both for animals to learn movements using reflexes and CPGs?

“This is basic research at the interface between robotics and biology. The robot model gives us answers to questions that biology alone cannot answer.”



Leave a Reply

Your email address will not be published.