What robots want to know

Robots are docile and quite limited beings. But that is going to change, said Ken Goldberg at the Digitising Europe Summit as he projected an image of a cartoon robot on the wall. A display on its chest read “You’re fired” in red letters, while the stoic expression on its face seems to say “I couldn’t care less”.

The audience, briefly surprised, smiles uncertainly – a reaction exactly as expected by Goldberg: “There’s no reason to be afraid that robots will take over and fire us,” he reassured his audience. “In the movies, robots can do a lot more than in real life.” However, Goldberg, Head of Automation Science Lab at University Berkeley, and his research colleagues are trying to change that – through “Cloud Robotics”. Their goal: learning robots.

Once a robot has expanded its knowledge, it can take over more complex tasks. So far, the machines have been set to perform specific movements and behaviour patterns. Anything out of the ordinary overwhelms them because they aren’t programmed for it. That isn’t a problem for robots that repeatedly perform the same processes, for example screwing together cars do the hoovering in the living room. They can be programmed not to bump into the table. But a home owner would get quite upset if his robot were to break a porcelain vase while tidying up because it gripped the delicate piece too hard. So how to remedy this? By more and even more specific programming?

Nope, unnecessary, says Goldberg. Instead of overloading the robot’s hard drive, it would make more sense to connect it to the Internet: There it can find all and any information it needs (and not least facts about the fragility of vases). “Robots can recognize, for example, 3D models of Coke cans and milk bottles in the cloud, compare them, and then grab them accordingly with the correct amount of pressure or caution”, says Goldberg. If a robot has several thousand models saved, it can classify similar objects automatically. Goldberg: “Then we’ll have actually achieved a self-learning robot.”

How Robots Learn

Once programmed, industrial robots can set welding points in the same place for months or years without needing much maintenance. But if a robot has to autonomously perform complex tasks the scope of which it can’t grasp, it becomes overwhelmed and uncertain. “Our job is to reduce this uncertainty as much as possible, to zeroise it,” said Goldberg, referring not only to the expertise of the machine, but also the manner in which it moves in space: it must react, adapting its performance again and again to the situation.

A robot perceives its environment through cameras, dynamometers or laser-based distance sensors. Its computer programs process sensor data and decide the next action. Then it grabs a vase, takes a step, or opens a door. It intervenes in its environment and perceives reactions via sensors. This cycle of perception, decision, and action is constantly traversed hundreds of times per second.

The novelty: The robot itself generates the sensor data from which it is to learn. It decides how to move, where to look and in which direction to walk. The result of this learning process changes the behaviour of the robot, for example it may move more efficiently or run more safely. As the researchers at the Max Planck Institute for Intelligent Systems in Tübingen have pointed out, this leads to a dilemma: in order to learn new things or to improve its performance, the robot has to test new behaviour, or intentionally deviate from already learned conduct. This means it could also learn “wrong” – for example by losing the ability to walk. “It’s a question of education,” according to the Future of Humanity Institute at Oxford University, which also deals with this issue.

What happens when people push the “red button” to stop bad behaviour? A robot could learn to undermine the effect of the off switch. Technology philosopher Nick Bostrom is already thinking about super intelligent robots that cannot be switched off.

Now Offices are Being Conquered

The learning robot has not yet left the research laboratories. But even its less inquisitive colleagues are now smart enough to move from factories to offices. Not only are workbenches being cleared, but also desks; machines have take over controlling, translations, and medical diagnoses, they sort libraries and design houses. The Mannheim Centre for European Economic Research (ZEW) has calculated that, in Germany, over five million jobs could easily be automated – most of them in offices.

The ZEW team based its work on a study by Michael Osborne and Carl Frey, in which they estimated that, in the US, machines ­ often robots – will take over 47 per cent of all jobs in the next two decades. “Anything that can be digitised and automated will be digitised and automated,” predicted Oxford economist Frey at the Digitising Europe Summit. Therefore, the majority of workers in transport and logistics professions as well as the majority of office workers will have to look for new jobs. Not all of them will find equivalent ones.

Even in factories, robots could take on new features: as soon as they will be allowed to leave their cage. Nowadays, production robots perform their welding, painting, and other daily work mainly behind bars ­ for safety reasons, so they won’t injure their human colleagues. “An industrial robot can be a dangerous fellow,” said Norbert Elkmann, director of Robotic Systems at the Fraunhofer Institute for Factory Operation and Automation (IFF) in Magdeburg. Such robots weigh up to five tons and exert enormous strength. Future models will be much lighter. Research is being conducted in IFF laboratories regarding collisions with robots that could be dangerous to humans and how they can be prevented. Elkmann’s goal: “Robots and people should work hand in hand”. And this should happen pretty soon. Experts believe that the share of human-robot systems in production sectors will increase sharply in the coming years.

Where Humans are Still Needed

Whether in factories or offices: robots will take over everything that they can do better than humans. They are faster, cheaper, and make fewer mistakes. And they are used in places where people would be overwhelmed or at risk:

• defusing bombs and mines;
• monitoring wind turbines;
• exploring the surface of Mars;
• working in extreme heat or cold;
• servicing tubes, shafts, and tunnels; and
• editing microchips in laboratories to a thousandth of a millimetre.

Well, where are humans still needed? Oxford economist Frey’s answer at the Digitising Europe Summit points the way: “In the areas of social intelligence and creativity human labour will continue to have a competitive advantage.” Blessed is he who understands how to solve complex problems.