It took nature millions of years to create intelligent, adaptive species. Researchers at Northwestern University in Illinois are using AI to evolve robots in minutes. The result is a robot that is agile, highly adaptive, and technically indestructible.
Robots are typically built with specific environments in mind. A warehouse robot glides efficiently across smooth concrete floors, while robot dogs can go up and down stairways without a hitch. However, these robots often have zero adaptability outside of predefined conditions. Even outdoor robots, which are typically able to navigate uneven terrain, may struggle when they encounter challenges they were not specially programmed to overcome.
The same robot that impressively sprints across an open field might as well be a toaster when it gets stuck in mud or, worse still, loses a limb.
The Northwestern researchers are exploring a different approach. Instead of designing robots for specific environments or trying to predict every possible scenario, they are experimenting with machines that can reconfigure themselves and adapt their movement depending on the situation.
Their result is the “legged metamachine,” a rugged modular robot that looks strange and moves even more strangely. Yet the researchers say it can literally be cut in half and still function.
The robot comprises multiple Lego-like modules that can assemble into various configurations. Each module is a functional robot containing a battery, motor, and computer. The design is relatively simple – a central ball and two adjoining arms (or legs) that can rotate along a single axis.
Alone, a module can independently roll, twist, and hop. Assembled, however, is when things start to get interesting. Communicating through the internal computers, the assembly can jump, crawl, roll, undulate, and perform a bunch of other movements that, frankly, are chaotic and hard to describe, yet quite effective. Basically, the robot does whatever it needs to do to get from point A to point B.
This is also the source of its “indestructibility.” If any module detaches from the assembly or sustains damage, the rest of the machine will find a way to keep moving, fulfilling its primary mission.
Perhaps the most interesting part of the project is who, or what, designed the robot – artificial intelligence.
As the creators of robots, humans are in charge of their evolution. So far, our thought process in this task usually involves adapting robots to match locomotors in nature or other existing locomotion technologies. As a result, we have invented multi-legged, tracked, wheeled, flying, swimming, and crawling robots. But what if these movements are not always the most ideal for robotic entities? The researchers turned to AI to answer the question.
Rather than designing the robot directly, the team provided AI with a set of building blocks, the modules, and a simple goal: find the most effective way to move. From there, the AI executed a process that closely resembled natural evolution.
“Evolution can reveal new designs that are different from or even beyond what humans were previously capable of imagining. So, we really wanted to study how and why it works. The best way, or at least the most fun way, is to evolve structures in realistic conditions,” explains Sam Kriegman, the study’s leader and a biorobotics expert.
Inside a computer simulation, the AI generated thousands of possible robot configurations. Each design was tested in countless extreme virtual environments. The most successful designs, those that moved farther or handled obstacles better, were kept, and the weaker ones discarded.
Generation after generation, the designs evolved until the AI eventually produced highly effective configurations, many of which humans would never have considered. The developers then physically assembled the final designs.
In outdoor tests, the resulting robot effectively moved across various rough terrains, including grass, gravel, and mud. For most robots, losing a single part renders the rest of the machine useless. On the other hand, the researchers’ robot was able to adapt and keep moving even after losing an entire leg.
“They [the robots] can survive being chopped in half or cut up into many pieces. When separated, every module within the metamachine can become an individual agent,” says Kriegman.
The researchers also claim this is the first evolved robot to step foot outside of a simulation and into the real world.
And just a quick sidebar ... Evolved robots, or evolutionary robotics, is a field that uses AI-based evolutionary algorithms to design machines by simulating natural selection. Instead of designing a robot directly, engineers design the rules of evolution. This approach essentially turns robot design into a form of digital evolution.
For fans of the movie “Big Hero Six,” the legged metamachine may seem oddly familiar. The movie features a robot concept in which swarms of tiny robotic modules combine to form larger structures that can reshape themselves at will. Each module comprises a ball and two arms, with autonomous control and features. The resemblance between that fictional idea and this real-world research is striking.
Granted, the real machine is obviously far less polished, but if the movie is anything to go by, the technology’s potential is huge.
For now, however, the technology is still in its earliest stages. At present, the machine has no outward-facing sensors. It cannot see obstacles or map its surroundings. It doesn’t even truly know where it is going. Most of its intelligence is focused inward. The robot can detect its own orientation (whether it is upside down or stuck) and the positions of the modules. On the other hand, it is slow, moves awkwardly, and lacks the sensing capabilities needed for real-world tasks.
Even the researchers themselves acknowledge that the robot is not yet particularly useful. But usefulness isn’t the primary goal, at least not yet. One of the aims of the project is to change the way engineers think about robots. Thanks to AI-powered evolutionary robotics, the researchers have achieved the most fundamental trait of any evolving species: survival.
A paper on the study was published in the journal Proceedings of the National Academy of Sciences.
Source: Northwestern University
