IF WE take the of prominent tech CEOs at face value, humanity could be risking extinction if artificial intelligence is left to its own devices. Large language models like those that power ChatGPT have shown the first sparks of intelligence, they say, and might lead to armageddon (although the specifics of exactly how have yet to be fleshed out). A few weeks ago, I found myself in an alternate AI reality – with zero mention of human extinction or analogies to nuclear war – when I attended the world’s largest robotics summit, the IEEE International Conference on Robotics and Automation (ICRA) in London.
It was a breath of fresh air to see how AI is being applied in an entirely different domain, though there are problems and societal risks with intelligent robots, too. How can robots adapt and learn in the real world without endangering people? Why is communication between fleets of robots such a hard and potentially dangerous problem? This area of AI research gets much less of the limelight than it deserves, given the outsize impact it could have on our future.
Advertisement
If you have seen robots on social media, like Boston Dynamics’ acrobatic android Atlas doing flips off a platform or the eerie being interviewed on This Morning, you might think robots already seem quite capable – but there is an element of smoke and mirrors here. Such robots have been trained to perform a narrow range of tasks in structured environments. When I go to the supermarket and see they have run out of eggs, I can adjust and try to find them in another shop, but a robot can’t yet do the same.
As Chelsea Finn at Stanford University in California told a packed crowd at ICRA, what we need is robots that can adapt to the almost infinite number of scenarios they might encounter. “Instead of programming robots or training them and deploying a fixed policy, maybe we can have robots adapt on the fly during deployment when they are in the real world,” she said.
Adapting and learning how to learn is a vast area of robotics research and we are beginning to see the first fruits of success in real-world scenarios.
Earlier this year, I reported on a project from Google’s robotics team that saw robots learn how to separate waste into recycling, compost and rubbish. This fleet of robots was taught using reinforcement learning, where a robot is shown a successful outcome and left to figure out how to achieve it by trial and error using a system of feedback. After lessons in robot classrooms and roaming the uniquely filled bins of Google’s offices for two years, the robots could sort the trash with 84 per cent accuracy.
A particularly interesting feature of this robotic fleet is that if one robot learns, they all learn. Although the robots are siloed in their physical bodies, they all share the same model of how to sort through a bin. This shared, or collective, intelligence is another thriving research field and something suited to robots. While humans rely on many forms of collective intelligence, we are limited by our communicative speed. Fleets of robots, however, can send lightning fast messages between one another.
“We’re at a time where robotics is becoming more pervasive, distributed and connected,” Amanda Prorok at the University of Cambridge told an audience at ICRA. Groups of robots that might communicate are everywhere: driverless cars, autonomous warehouse robots, swarms of drones. But for individual robots to use the data they are gathering and then communicate it with each other intelligently is hard. Errors and uncertainties in communications can quickly compound and spiral out of control, and the real, noisy world is full of these errors.
Prorok demonstrated how even a simple problem – like ensuring the flight paths for a few drones aren’t disturbed by the gusts of wind from each other’s rotors – requires complex, cutting-edge AI. Scaling up these sorts of problems to much larger systems will involve a lot of work, and the downsides of getting it wrong are potentially large. If you are driving somewhere and your car ends up sandwiched between two moving driverless ones, it might make sense for those two vehicles to communicate about what is happening, but those communications must be fair and avoid endangering you.
Despite these potential pitfalls, I couldn’t help but feel excited at the future as it was presented at ICRA. There is a chance that much of this will fail – fully automated cars have been just around the corner for over a decade – or descend into a dystopia, as we have seen with drone warfare in Ukraine. But I was heartened by the brainpower and resources being poured into delivering on robotics’ positive promise.
New Scientist video
Watch Alex Wilkins try out robot avatars at the ICRA robotics conference
Alex’s week
What I’m reading
Seeing Like a State by James C. Scott, a look at how things go wrong when state schemes to improve society fail.
What I’m watching
Beef, a Netflix series about escalating acts of revenge between two angry drivers whose lives become intertwined.
What I’m working on
There’s no time like the summer to finally get the garden into shape – no easy task for a hayfever sufferer like me.
