PADDING through a deserted office in Waltham, Massachusetts, a quadrupedal creature stops in front of a pair of heavy doors. Resting on its haunches for a moment, as though contemplating the obstacle, it turns and seems to summon a friend. The two could almost be Dalmatians. Until, that is, one of them turns what appears to be its head into an articulated arm that grabs the handle, twists it and pulls open the door.
It makes for a weirdly enthralling scene. So enthralling, in fact, that millions of people have watched the YouTube video of these SpotMini robots doing their thing. They are astonishing pieces of engineering, capable of feats of locomotion and dexterity that took evolution millions of years to perfect. But they aren’t quite as capable as they seem. For all their physical prowess, the intelligence of these mechanical creatures is sorely limited. They can walk with an unnervingly animalistic gait and now they can open doors, but what they can’t do is, well, pretty much everything else.
That may all be about to change. Techniques from the best AIs are already being adapted to endow grasping robots with the ability to learn new skills for themselves. The robots in question are still clunky, not to mention stationary. But if this sort of software can be integrated into agile robots like the SpotMini, things will start to get interesting. From doing the household chores to saving you from a burning building, robots that move and learn like us could finally transform our lives.
Advertisement
For creatures as brilliant as us, it is easy to forget that moving is hard. When you walk, or even just stand still, your brain is constantly telling your muscles to make thousands of tiny adjustments. Nobody knows that better than roboticists. For the most part, their attempts to build machines that can do some of the things animals do have been hopeless. A quick glance at the footage from the world’s foremost robotics competition proves that. Basically, most of the time, even the most advanced robots fall over.
[video_player id=”htcFBzMu” access_level=”everyone”]
One man who has felt this particularly keenly is Marc Raibert. As the founder of Boston Dynamics, he is responsible for the spectacular performance of SpotMini (featured in the video above) and its humanoid companion Atlas. Raibert began a career in academia in the 1980s, aiming to build robots that mimicked the movements of humans and other animals. But according to Joanna Bryson, then a graduate student at the Massachusetts Institute of Technology (MIT), it was being embarrassed in front of Sean Connery that radically transformed the scope of his ambitions.
In 1992, Connery was making the film Rising Sun in California. The director wanted a couple of scenes to feature robots, and Raibert’s group was happy to oblige with its two-legged machines designed to hop about and maintain their balance. The trouble was, the robots constantly broke down and generally wound up the film crew. As a result, Raibert vowed to make robust, reliable robots on a commercial scale. Within the year, he had set up Boston Dynamics to build machines with “mobility, dexterity, perception and a kind of intelligence that holds it all together” – and he has gone a long way to making good on his promise.
Over the past three years, Raibert’s robots have cornered the market in viral videos of machines doing amazing things. For SpotMini, the list of accomplishments captured on film includes not only opening inward-swinging doors, but also getting back up after slipping on a banana skin, doing the Running Man dance and navigating construction sites, which involved negotiating narrow staircases and various obstacles along the way. For Atlas, the highlights reel includes running through the woods, doing backflips and picking up parcels. Little wonder views of robots in action on the number in the hundreds of million.
“People assume if it walks like a duck it is a duck. It is not. It is a robot”
“You can’t fail to be impressed by what they do,” says Bryson, now an AI researcher at the University of Bath, UK. “These things really work. When you go to Boston Dynamics, there are robots running around. They do do all the stuff you see in the videos.”
Raibert freely admits that the footage shows the robots on a good day. He has said, for instance, that the clip where Atlas runs up three big steps as if doing parkour took 22 attempts. And in his TED talk from 2017, he showed a clip in which Atlas picks up a box, misses the shelf it was aiming for and pulls over a cart as it falls flat on what passes for its face. It was a sobering display of robotic ineptitude.
None of this takes away from what Boston Dynamics has achieved. For Ioannis Havoutis at the Oxford Robotics Institute, UK, the most impressive thing is how well these robots maintain their balance. “As a control problem, this is particularly hard,” he says. SpotMini also seems to have a good deal of autonomy in its navigation, he adds, in the sense that you can command it to go from point A to B and it will find its way without a human pilot.
Fear not, though, because that is about as far as their intelligence goes. Watching their performances, it is easy to assume that these robots are autonomous. The reality is that there is almost always a person controlling them via a laptop and an Xbox controller, sending instructions such as “go forward”, “turn around” or “open the door”. Moreover, when the SpotMini does open the door, it is following a script: painstakingly hand-coded instructions that tell it how to reach for and turn that type of handle.
This is something people tend to miss. Because they look vaguely like animals and their movements seem quite natural, we think of them as animals and then assume they have all the other abilities those animals have. For anyone familiar with the scene in Jurassic Park where velociraptors struggle to open a kitchen door, the SpotMini clips can be particularly unsettling.
“People assume that if it walks like a duck it is a duck,” says Bryson. “It is not. It is a robot.” Just because SpotMini can follow a script to open a door, for example, doesn’t mean it can do other dexterous things, never mind chase you down and rip you limb from limb.
“These robots are really good at the specific tasks they were designed for,” says Havoutis. “But there is no general AI there.” Or, as Raibert himself has put it, the world’s most advanced robots “can’t do almost everything”.
In that light, it is tempting to raise a sceptical eyebrow at SpotMini’s achievements. After all, YouTube fame hasn’t translated into revenue for Raibert’s company. Google snapped it up in 2013 and it was sold on to Japan’s Softbank last year, a move that many saw as an indication the US tech giant couldn’t see how these robots would ever make money. And yet the engineers at Boston Dynamics, and indeed elsewhere, have clear ideas about where legged robots will find employment.
Where humans fear to tread
The first are likely to be places where humans would rather not work, because they are too dirty, lonely or dangerous, ranging from sewers to oil rigs. Havoutis and his colleagues have already used their ANYmal robot, which looks similar to SpotMini, to carry out inspections on a mock oil rig outside Oxford. It won’t be long, he hopes, before these robots can intervene directly, by turning a valve or pressing a button, so long as they are teleoperated by a human.
In the longer term, such robots might also find use in disaster response. “Our ultimate goal, in the next 20 years, is to build a robot that can save a life,” says Sangbae Kim at MIT, who works on another agile quadrupedal robot called Cheetah III. “I remember talking to a fire chief who told me he is sometimes faced with an impossible decision: he doesn’t know if there is anyone inside a burning building, so does he send in a firefighter and risk their life? With a robot that can find someone and bring them out, there would be no such questions.”
And then, of course, who doesn’t love the idea of a robot butler, a machine so reliably perceptive and dexterous that it would tidy your house and whip up an apple crumble, just in time for when you get home from work? The trouble is, robots that can operate at that level require a crucial component that is currently in short supply: general intelligence.
Right now, even the world’s most advanced robots just aren’t very bright. Even if we could endow an Atlas with hardware for the fine manipulation involved in simple chores, like picking up an apple and putting in a bowl, programmers would have to write reams of code to make it possible. And if something unexpected happened, the robot wouldn’t have a clue how to react. “When it comes to high-level intelligence, we’re not even nearly ready,” says Kim.
What robots need now is the ability to learn new skills for themselves (see “The importance of a body”). The good news is that AI researchers are developing nifty ways to help them do exactly that. In Sergey Levine’s lab at the University of California, Berkeley, for instance, a clunky looking humanoid robot called PR2 can learn to pick up an apple and place it in a bowl – even if it has never seen an apple before – just by watching a human do it.
It isn’t exactly brain surgery. Nor does it make for footage that will take the internet by storm. But by demonstrating techniques that let robots autonomously acquire new skills, Levine and his colleagues are taking the first big steps towards making robots that have what it takes to do the household chores – and much more besides.
“If you want a robot to do anything useful, you need it to be able to adapt on the fly to new situations,” says Chelsea Finn, a former PhD student in Levine’s lab and now at . “That’s what we’re trying to do.”
One of the big challenges is to get the software to recognise when it is succeeding or failing. For tasks where massive data sets of labelled images exist, that learning process is comparatively straightforward. Take the chess program AlphaZero, developed by Google sister firm DeepMind in 2017. That took a mere 9 hours to progress from total novice to become the strongest chess-playing entity in history. In that case, all possible moves and outcomes can be perfectly calculated, allowing the AI to continuously play against itself, generating an unlimited supply of data in the process.
If they only had a brain
The real world is less clear-cut. Even so, researchers like Finn and Levine have two strategies. The first is known as imitation learning: showing the robot how to do something by either teleoperating it or having it watch videos of a human doing something. The other is reinforcement learning and, in some cases, it requires no supervision whatsoever. Last year, for instance, Levine and Finn demonstrated a robotic arm teaching itself to do various tasks, from picking and placing apples to folding shorts and – somewhat idiosyncratically – wrapping forks in towels, based entirely on data collected from its camera as it fumbled around with these objects. “Basically, the robot learned to predict what happens if it performs an action,” says Finn. “Then we give it some goal and it figures out what actions it should take to achieve that.”
It is much like the way children learn by playing and, although the tasks themselves aren’t complicated, the sheer number of things it can do with this one algorithm is impressive. “Traditionally with machine learning you train one model to do one thing – translate French into English, say – and you always start from scratch,” says Finn, “but here you can build a single model and use it for many different tasks. That is very powerful.” So powerful, in fact, that another of Finn’s former supervisors at Berkeley, Pieter Abbeel, has launched a company called covariant.ai to push these techniques out to warehouses and factories.
For the time being, most robots learning for themselves are stationary, cumbersome and confined to the lab. Even so, Levine says there is no reason these techniques can’t eventually be incorporated into legged robots. In fact, the team behind ANYmal has recently shown that its robot is capable of reinforcement learning, teaching itself to run faster and pick itself up from a fall more efficiently with no human intervention. For Levine, this is the future. “I have no doubt that, in the long run, robots that are constantly improving themselves will be preferred.”
It seems safe to assume that legged robots will, at some point, be able to teach themselves new tricks. So is there any reason to be scared? If dexterous robots can learn for themselves, what would stop them from picking up other skills, using those claw arms not just to deliver parcels but to throttle unsuspecting humans on the doorstep?
For Levine, the fear of AIs somehow going rogue is misplaced. A bigger risk, perhaps, is the military use of robots, or of existing AIs somehow being subverted. “Machine-learning systems are no different to other computer systems,” he says, “in terms of being vulnerable to attack from malicious hackers.”
For the time being at least, our primary concern about autonomous robots should be that they are too stupid rather than too smart. Imagine a clumsy robot attempting to rescue a family from a burning building, or an inflexible AI controlling heavy machinery in unfamiliar circumstances. A world of legged, self-learning robots let loose outside the lab is still some way off, but when that future does arrive, we need them to be as intelligent as possible. When opportunity knocks, let’s see if they can open the door without pulling it off its hinges.
The importance of a body
The fusion of artificial mind with matter is already paving the way for autonomous robots that could do your household chores and it may yet do a lot more than that. For some experts, this convergence of robotics and artificial intelligence could mark a turning point in the quest for machines with something approaching human-level intelligence.
The idea is that if you want AI to fulfil its potential, it needs to interact with the real world, like humans do. Only then can we expect it to undergo something akin to the process that gave rise to human intelligence.
“Embodiment is a critical part of how humans and many animals learn, because it allows you to build and test hypotheses,” says Chelsea Finn, . As babies, we learn by exploring the world, throwing toys around or spilling water. For Abhinav Gupta at Carnegie Mellon University in Pittsburgh, Pennsylvania, and part of Facebook’s AI team, this is the only way AIs can achieve the kind of common-sense knowledge about how the physical world works that we take for granted.
“Current AI is built on solving specific tasks with large amounts of data and supervision,” says Gupta. “To build AI that can solve general-purpose tasks, that can reason in domains with very little supervision, we need AIs to learn predictive models and causal, common-sense knowledge. Embodiment allows us to do that.”
Not everybody has such high hopes. “Many animals have very sophisticated ways of interacting with the environment and have nothing like human intelligence,” says evolutionary psychologist David Geary at the University of Missouri. For robots to reach our intellectual heights, he believes, some process akin to evolution would be required, specifically designed to weed out robots with poorer conceptual abilities. In such a case, says Geary, robots might even shed light on the evolution of human cognition.
