Researchers at Germany's Max Planck Institute for Intelligent Systems (MPI-IS) have made a robot dog that can learn to walk on all fours in about an hour. In the same way that newborn animals quickly learn to stand on their own feet after birth, a robot dog taught itself to walk in just one hour.

This research might look like the start of a robot rebellion, but it was part of an experiment to see how animals learn to walk by falling. For example, when human babies try to walk for the first time, they often trip and fall. These "mistakes," on the other hand, help us adjust our leg muscles and nervous system so that we can walk normally in the end.

Morti

Morti, the mechanical dog, is programmed with an algorithm that works like the central nervous system. Instead of following complicated programming, Morti acts on instinct and corrects his walking mistakes as he goes. As a result, it acts reflexively, correcting walking mistakes in real-time instead of following complicated programming.

The robot learned to walk by constantly looking at feedback from sensors that told it if its feet were hitting the ground correctly and by keeping an eye on its overall balance and how smoothly it moved. If it fell, the learning algorithm changed things like how far back and forth the legs moved, how fast the legs moved, and how long the legs were on the ground. 

Morti struggled at first to maintain balance and stand up straight while being placed on a treadmill, stumbling and sliding along as it went. However, the robot's four legs could walk smoothly after an hour.

Research overview

In the journal Nature Machine Intelligence, they wrote a paper about their study. In their article, the researchers looked at how a walking system with limited control and sensor bandwidth could learn to use the intelligence in its leg mechanics. Robotic systems are often judged by how well they use energy and how fast they move.

Image source: Nature Machine Intelligence

Morti is a four-legged robot. (a) A picture of Morti. (b) A drawing of Morti on top of the treadmill.

In this case, the researchers suggested an extra step focusing on how passive mechanical structures and neural control work together. By separating feedback by how long it takes to happen, they could reduce disturbances in the short term and figure out how much the control patterns and natural dynamics didn't match up. Finally, the researchers worked to improve the system's long-term performance and fit the controller to the system's mechanical parts.

Even though the researchers researched a robot stand-in, the results could shed new light on how learning might happen in biological systems with limited neural resources and little feedback. Of course, matching is probably not the only thing that makes animals learn. Still, their study suggests that a quantitative measure for "long-term learning from failure" could be affected by the goal of maximizing the synergy between locomotion control and the robot's or animal's mechanical walking system. Furthermore, in contrast to task-specific cost functions like speed or energy efficiency, their matching approach gives them a reason to use as much of their natural intelligence as possible.

Conclusion

Animals can move quickly and easily with less control effort and less wasted energy by using the flexibility of their muscles and tendons. But no one knows how biological controllers of locomotion learn to use intelligence in their leg mechanics. 

Here, the researchers use the idea of short-term elasticity and long-term plasticity to show how we can match control patterns and mechanics. Researchers make Morti, a robot with passive elastic legs, by taking ideas from animals. Morti, a four-legged robot, is controlled by a bio-inspired closed-loop central pattern generator that uses sparse contact feedback to adapt to short-term changes. By reducing the amount of corrective feedback over time, Morti learns to match the controller to how it works and walks in an hour. In addition, Morti improves its energy efficiency by 42% by taking advantage of the good things about how it works. Significantly no reduction in the cost function during this process. Furthermore, we could utilize the idea to create energy-efficient robots that do not spend energy on sophisticated processing.

Image source: Unsplash