Roboticists could learn a thing or two from insects if they’re looking to build tiny AI machines capable of moving, planning, and cooperating with one another.
The six-legged creatures are the largest and most diverse multi-cellular organisms on Earth. They have evolved to live in all sorts of environments and exhibit different types of behaviors to survive and there are insects that fly, crawl, and swim.
Insects are surprisingly intelligent and energy efficient given the size of their small brains and bodies. These are traits that small simple robots should have if they are to be useful in the real world, a group of researchers posited in a paper published in Science Robotics on Wednesday.
“We argue that inspiration from insect intelligence represents an important alternative route to achieving artificial intelligence in small, mobile robots,” they wrote. “If we succeed in harnessing insect-inspired AI, small robots will be able to tackle difficult tasks while staying within their limited computational and memory budget.”
Roboticists are already building bug-like bots. Guido de Croon, first author of the study and a professor at the faculty of Aerospace Engineering at TU Delft University in the Netherlands, helped develop a swarm of tiny drones designed to detect gas leaks in buildings. Elsewhere, researchers at the University of Washington in the US built the first wireless flying robot, complete with a pair of wings, not much heavier than a toothpick yet capable of taking off and landing.
They may not be so impressive compared to larger, more complex machines, but their tiny size and simple electronics make them cheap and potentially useful for applications such as search and rescue, surveillance, or even pollination. Significant challenges remain in building these machines, however, even with the advance of newfangled AI algorithms that have advanced computer vision, planning, and navigation due to hardware and size constraints.
“Many deep neural networks that are being developed in AI are in principle interesting but will not be able to run on small robots yet,” de Croon told us.
“For example, there are neural networks that estimate visual motion or recognize objects. Embedded computers made for running deep neural networks are typically on the heavy side and quite power consuming.” Even the smallest GPUs, designed for embedded electronics and capable of running these AI models, are right now too heavy and power hungry for small flying robots that have to be as light as possible.
“While a popular embedded processor for deep nets, the Nvidia TX 2 weighs 85 grams and consumes 7.5 Watts. Honestly, even for slightly larger and heavier drones, the relative weight and power of deep net processors should go down,” he added.
There are hardware alternatives from Croon and his colleagues believe are promising – microcontrollers and other chips for tiny embedded systems are gaining the necessary oomph for performing ML tasks – while more futuristic neuromorphic processors are better suited to running machine learning algorithms more efficiently.
Intel’s neuromorphic chip, Loihi, for example, powered a spiking neural network model to control a flying robot. The end goal, however, isn’t necessarily running today’s complex software on new hardware, the researchers argued. Real progress will come in developing novel algorithms and models capable of running on energy-efficient hardware incorporated in machines that can replicate insect intelligence.
“The main property of insect intelligence is its parsimony, that is, the way in which insects use minimalistic yet robust solutions to achieve successful behavior in complex, dynamic, and sometimes hostile environments,” according to the paper.
De Croon told The Register it was “important to read the biological studies by entomologists” to find inspiration. “Interestingly, though, it is not a one-way street: When trying to design robotic systems for performing tasks done by insects, we often run into problems that are not always evident when directly studying the animals. This in turn may lead to novel insights in biology, that can then be studied by working together with entomologists,” he said.
When trying to mimic the motion of fruit flies in one experiment, his team was able to study the mechanism of how they flapped their wings during escape maneuvers.
Mimicking insects mechanically will also advance other areas of robotics. “Insect-like intelligence is also relevant to many other types of robots, as it brings robustness while taking as little resources as possible,” he concluded. ®