Researchers at the University of California, Berkeley, have used AI to teach a robot dog to walk in the real world.
The AI method, called reinforcement learning, trains algorithms by giving them rewards after they perform desired actions.
The robot dog, named Morti, taught itself to walk within an hour of its first step. Its process mirrored that of young animals in the wild, with it stumbling until it learned to balance.
Within an hour, it could walk on a treadmill.
Morti was built to help study how animals learn to walk, with the researchers taking measurements of each of its limbs’ forces, torques and muscle power. The experiment would therefore be much harder to replicate with a living animal.
Unlike Morti, who runs on four small nodes, most four-legged robots rely on an intricate controller to manage their movement.
“If we talk about industrial application, battery life is always a bottleneck,” said Felix Ruppert, a PhD student and one of the authors of the study.
“So if the computer uses less energy, then a smaller, less energy-hungry computer can be used.
“The point of our approach is not to replace what is currently used, but to provide additional aspects that could be synergised with the current state of the art technology, to create the next generations of walking robots.
“What drives learning to walk? What is the best placement of muscles and how long should legs and their segments be?
“And, more broadly, why are animals so good at locomotion all the neural delays they have – and why have we not been able to reproduce that performance in our robots?
“Bio-inspired robots are a great tool to understand biomechanics and its unanswered questions.”
Walking, blinking and breathing are called rhythmic tasks and aren’t co-ordinated in the brain, but are controlled by a central pattern generator (CPG), which is formed by a network of neurones.
Humans’ CPG for walking is found in peoples’ spinal cord, which means when humans trip, they don’t immediately stop walking, because the spinal CPG controls the legs’ reflexes without needing input from the brain.
This means an animal’s ability to walk is also stored in the spine, which lead the roboticists on the study to develop an algorithm that functioned like a computerised spinal cord for Morti.
