Many AI robots appear intelligent as long as they operate in controlled environments. Neuroscientist and robotics researcher Elisa Donati explains why robots with real-world readiness require more than just fast software.
Robots already perform a huge range of tasks: they work on factory floors, deliver parcels and clean our apartments. However, many systems become surprisingly clumsy once they are expected to navigate an unpredictable real-world environment. This is why developing AI robots requires more than a narrow focus on artificial intelligence - we also need to consider the interplay between brain, body and environment. Intelligence, it turns out, is not contained in software alone.
When a robot has an intelligent body, its "brain" can afford to relax a little. One good example can be found in the animal kingdom: many animals are far less intelligent than humans and possess only a rudimentary nervous system, yet they are capable of orienting themselves, navigating around obstacles and interacting with their environment. Part of this behaviour emerges directly from the interplay between body and environment.
Applying this principle to robotics makes it clear that a robot's physical design is just as important as the AI that powers it. Soft robots illustrate this idea particularly well, since their pliable bodies make certain tasks much easier. If a robot's hand needs to carry a delicate porcini mushroom, for example, a rigid robot would be wrong for the job, as it would need extreme precision to carry out the task. A soft robot, by contrast, yields gently on contact, making the task much simpler. Because its soft structure absorbs some of the contact forces, the robot's AI-powered brain needs to exert far less control.
The expert
external page Elisa Donati heads a research group at the external page Institute of Neuroinformatics at ETH Zurich and the University of Zurich and is a titular professor at the University of Zurich. The neuroscientist and robotics researcher investigates how neuromorphic AI systems - computers inspired by the architecture of the brain - can be integrated with robotics to make robot-environment interactions faster, more adaptive and more energy efficient.
If intelligence isn't merely a property of software but something that emerges from the interplay of body, brain and environment, then evaluating robots purely on speed or precision is no longer sufficient.
Everyday life is messy and full of unpredictable situations. To offer real assistance in this kind of environment, AI robots need a certain degree of flexibility when responding to change - for example, when they encounter obstacles, different types of surfaces or unfamiliar tasks.
This is why we also need to measure how quickly robots adapt to new situations, how robust they are under real-world conditions and how efficiently they use energy. How they interact with their environment is equally important: Do they cause any damage? Can they work safely with humans?
What we need, then, is adaptability over perfection. The future doesn't necessarily belong to the cleverest systems, but rather to the most flexible ones. However, a robot's capacity for flexible responses is only as great as its onboard computing power. There is no room for a supercomputer inside of a multi-purpose everyday robot.
My suggestion for making robots suitable for everyday use is to combine small, specialised control networks with an intelligent body, as published in a recent external page study . When looking at the natural world, we see that there aren't any central systems governing every movement and reaction. Reflexes and motor sequences often arise from local neural circuits that handle specific tasks with remarkable efficiency.
Some aspects of intelligence emerge directly from the interplay between body and environment. This holds enormous potential for robotics. AI-powered robots aren't yet perfect all-rounders, and perhaps they don't need to be. Far more important is how well they cope with our messy world and how we can learn to live alongside them.
