Why improving robot design is essential to achieving true intelligence

Thanks to artificial intelligence, robots can already perform many tasks that would otherwise require humans. In this interview, Edoardo Milana, a junior professor of soft machines in the Department of Microsystems Engineering at the University of Freiburg, explains how improved design and innovative mechanics are broadening the range of applications for these machines.

Why is there a need for an alternative to conventional robots?

Robots can already perform amazing things with the help of artificial intelligence and machine learning. However, all this intelligence is concentrated in the software—the brain—and no comparable focus is placed on the mechanical design—the body. As such, robots are pretty much like puppets. Software is used to try to exert full control over all the body’s movements.

This approach requires the hardware to be very simple from a mechanical point of view and easier to operate using digital microcontrollers. Depending on the area of application, this may be sufficient and even needed to meet the precision and high force requirements. However, when we look at motion efficiency and agility, the performance of robots falls far short of that of living beings.

Nevertheless, there are already robots that imitate animals, such as dogs and cats

These quadrupeds—and even humanoids—are impressive engineering masterpieces, however, they cannot compete with real animals in terms of motion agility. They also consume a lot of energy to move, whereas animals and humans can perform much more complex movements using far less energy. A quadruped walking at normal pace consumes roughly 300 Watt on average to drive its 12 motors, the robot’s “muscles,” while a dog consumes 30 Watt to activate hundreds of muscles.

This is possible because, in nature, movements strongly rely on the mechanical properties of the body, which passively and actively adapts to the external forces exerted by the environment, harnessing the compliance of biological materials. Beyond digital control, the focus in robotics should also be on implementing intelligence, or “embodied intelligence,” into the design of the robot. This would free up computing capacity and energy currently used for low-level motion control for the high-level logical operations of the robot, such as reasoning, planning and perceiving.

The concept of embodied intelligence originates from the fields of philosophy and psychology. But what does it mean for you as an engineer who develops robots?

For me, the interesting thing is that the theory behind it can be applied not only to biological beings, but to robots, too. The basic idea is that physical interaction between the body and the environment shapes intelligent behavior. It’s not just about having a body controlled by the mind—this control lies partly in the body itself and in the way it interacts with the mind.

In robotics, this means that if we want a truly intelligent robot, we can’t just build a body consisting of two or three metal bars and a few joints, then put a very intelligent computer inside it. If that were the case, we would already have robots with completely different capabilities.

What could such intelligent robots look like instead?

I am researching soft robots made of soft materials, which could be considered as inspired by primitive and aquatic biological organisms. There are already robots in this field whose control is based entirely on physical principles, and that do not require digital microcontrollers. They utilize the non-linear physical properties of soft materials to generate the control signals that drive the robot.

Together with researchers from Stuttgart, the Netherlands, and Belgium, I have written an article published in Science Robotics presenting such soft robots, which introduces a new concept: The concept of “physical control.” We have identified three particular control mechanisms for such soft robots.

One interesting example is robots with self-oscillating valves. When air pressure is added, the valves open and close again, increasing and then releasing air pressure. This transmits a rhythmic air pressure signal through the system, controlling the movement of the individual robot parts.

In the future, we will need to find a compromise: We won’t be able to manage without software and microcontrollers in robotics, but we can achieve a lot through better robot body design.