Pitting robots against each other shows the promise and many of the pitfalls of autonomous technology
THE BRIGHT red robot lay sprawled across the threshold, seemingly unable to go on. Chimp, built by Carnegie Mellon University’s National Robotics Engineering Centre (NREC), had successfully driven a car down a short simple slalom track. It had climbed out of the car, though not without some difficulty, headed over to a door and opened it. But after starting to move through the door Chimp came to a sudden halt; with a lot of inertia (it weighs 200kg) and its centre of gravity a little too far forward, that caused it to topple.
It was not the first to fall. The DARPA Robotic Challenge (DRC) was a three-year competition in which research teams from a number of countries tried to get robots to do things that might be required after a disaster like that at Japan’s Fukushima nuclear plant in 2011. Its finals, held in Pomona, California, on June 5th and 6th, saw robots fall every which way. They fell on their faces. They fell on their backs. They toppled like toddlers, they folded like cheap suits, they went down like tonnes of bricks. There were dents, breakages, spasms and even bleeding (or, at least, the catastrophic loss of hydraulic fluids). What there was not was any getting back up again. Except, that is, for Chimp (pictured below).
The NREC team operating Chimp had no specific plans for the position the robot found itself in. What was more, by sticking its head through the doorway Chimp had entered the part of the DRC course where communications between the robots and their controllers were deliberately degraded to increase the need for the robot to function autonomously. But Chimp’s programming put various ways of wriggling at its disposal, and it eventually managed to stretch itself out into a position in which it knew how to get back up. The watching crowd went wild.
Even so, Chimp did not win the finals. That honour went to Hubo, a smaller robot developed at the Korea Advanced Institute of Science and Technology (KAIST), which accordingly took away the $2m first-place prize from DARPA, the Pentagon-run technology hothouse which came up with the DRC. The competition was DARPA’s way of pushing forward the development of robots capable of working in an environment designed for humans but too hazardous for them—hence the requirement that they drive cars, open doors, use power tools, plug and unplug electrical cables, climb stairs and so on, all under human supervision but not always or necessarily under their direct control.
Chimp on a roll
It doesn’t sound hard. But robots are stupid; those not programmed for precise tasks tend to be clumsy, too, and on occasion perverse. The machines’ failings are oddly, and touchingly, moving to spectators like those in Pomona; they are deeply frustrating to their designers. When the DRC began in 2012 there were no machines able to do all the tasks imagined, let alone do them one after another. One measure of the programme’s success is that by its end three contestants managed to complete the final sequence of tasks in under an hour: Hubo, Running Man (pictured above in a less successful attempt, a robot run by the Institute for Human and Machine Cognition—IHMC—in Florida), and Chimp—which, having taken third place, walked off with a cheque for $500,000.
Or, rather, trundled off. Though Chimp has four limbs, like almost all the robots signed up for the finals, unlike most of the others it does not walk. Instead it uses caterpillar tracks on what one might think of as its forearms and feet. It can move around in an upright position that leaves two upper limbs free for doing things like twisting valves or cutting holes in walls while the lower limbs handle the locomotion—a stance that can indeed look vaguely simian. When it needs to get across rugged terrain, it can crouch on all fours, presenting a more turtle-y aspect. That versatility turned out to be quite a help.
Hubo doing a turn
The winner, Hubo, is more conventionally humanoid and can walk quite well. But it, too, enjoys inhuman options when it comes to mobility; wheels in its feet and knees allow it to scoot around with its legs tucked up underneath it, providing a quicker and more stable alternative to walking. In fact, it was no coincidence that two of the three winners used multiple means of movement. Of the 23 robots attending the trials, 18 walked from A to B, while six enjoyed an ability to trundle. Five of the six all ended up in the top eight places.
That walking is hard did not come as news to the DRC’s designers. Indeed, it was one of the points of the exercise. Developing a highly capable humanoid robot from scratch is beyond the capabilities of most university labs that might have good ideas about how to improve one. Believing that walking autonomous robots are a potentially useful, even transformative technology, DARPA designed the DRC to lower the barriers to entry, encouraging researchers who had not previously grappled with its problems to get involved.
To this end the agency had Boston Dynamics, a company that has built various walking and running robots for military programmes, create a two-legged, two-armed, almost-two-metre-tall robot called Atlas. A preselection round, held in mid-2013, saw over 100 teams develop software to control a simulated Atlas in a virtual environment. The groups that did best were given a real Atlas to work with, as well as money to run their teams. The development costs of some of the other competitors, such as Chimp, were also subsidised by DARPA. In total, the competition cost close to $100m.
Running Man finds his feet
Running Man, the Atlas robot with which IHMC won second place, showed what the platform was capable of—though after completing its winning round it did rather let the side down by falling over as it struck a sequence of victorious poses. Jerry Pratt, who led the IHMC team, argues convincingly that, in principle, walking has huge advantages—a human can quite easily make progress along a discontinuous track no wider than a single foot, taking in its stride obstacles big enough to pose a problem to the wheels of anything short of a monster truck. But on the evidence of the DRC, the software and hardware needed to match that ability remain far off. For the time being, a robot designed for responding to DRC-style disasters looks likely to need an alternative to legs.
Getting a handle on how long that time being might be was another of the points of the DRC. Gill Pratt, who ran the programme at DARPA (and is no relation to IHMC’s Mr Pratt, though he did supervise his doctoral research) saw it as a way not just to stimulate progress in the field but also to gauge how quickly such progress could be made. Everyone involved in the DRC remembers the startling improvement between the first of DARPA’s “Grand Challenges” for autonomous vehicles—which, in 2004, saw the winning car travel just 11.8km of a 240km route—and the second, in 2005, in which five teams went the whole distance. That demonstration of rapidly expanding capabilities played a role in convincing people, such as the bosses of Google, that self-driving cars were a practical possibility in the not-too-distant future.
The progress between the first real-world DRC trials, in late 2013, and the finals this month was less spectacular. Humanoid robots are not yet at the ready-for-take-off point autonomous cars were at ten years ago. The teams using the Atlases knew that less than two years of working with their charges gave them time to implement little more than a simple ability to walk—one expert says that developing robust locomotion from the ground up and debugging it is more like a five-year job. There were also limitations with the hardware. Atlas’s arms were not strong enough to lift its 150kg bulk back up if it fell down.
Chris Atkeson of Carnegie Mellon, who joined an Atlas team led by Worcester Polytechnic Institute in Massachusetts, says their group decided that in the time available there was only a limited amount they could do; they chose to focus on software and operational approaches which would stop their robot from falling down. Much to Dr Atkeson’s delight, theirs was the only non-trundling robot to stay upright throughout—though on neither day of the finals did it complete all the tasks.
A decade or so will make all the difference. One of the reasons the IHMC team did the best of the seven Atlases was its greater experience with bipeds. The same was true of the Hubo team—Jun Ho Oh, its leader, designed the first Hubo more than a decade ago. Experience, though, is not everything. Japanese researchers have more experience with humanoid robots than anyone else on Earth, but their teams fared very poorly. Dr Pratt at DARPA says this was because they fell between two stools. Some had tried to design brand new robots for the event and not got far enough with the software; others had instead depended on older robots that were not up to the tasks. One observer noted that the results might actually do Japanese robot labs a world of good; being trounced by a South Korean team would fire up national pride and open up chequebooks. It might also encourage the idea that Japanese robotics, something of a world of its own, has lessons to learn from elsewhere.
From science fiction to reality: A graphic guide to some of the most advanced robots
Ironically, in the DRC’s 2013 trials a Japanese robot, SCHAFT, swept all before it. If it had continued in the competition it would have entered the finals as an odds-on favourite. But the startup company behind SCHAFT, along with half a dozen other robot companies—including Boston Dynamics, the makers of Atlas—was bought by Google just before the 2013 trials. Team SCHAFT subsequently dropped out of the competition. What it is doing now is, like everything associated with Google’s robot developments, a mystery even to well-connected Silicon Valley robot experts, according to Brian Gerkey, who runs the Open Source Robotics Foundation there.
The aim of the DRC was not just to get more researchers in America interested in humanoid robotics but also to help them form a community of shared interests and experience. By providing a guide to who is best at it, the process also showed who to acquire. Larry Page of Google, Elon Musk of SpaceX and Tesla, and Travis Kalanick of Uber all visited the finals. In recent months Uber has stripped a great swathe of talent from the NREC—including the head of the centre and leader of the Chimp team, Tony Stentz. Dr Gerkey says he recently heard an official at another government agency which funds robot research voice fears that having so much talent sequestered in projects that the nascent community cannot track or learn from might leave a “research desert” behind.
Dr Pratt of DARPA acknowledges the worry, but says there is no real danger as long as the projects going on in the dark come out into the light eventually. With the DRC complete, he himself is now leaving DARPA for a new challenge—though he is not at liberty to say what the challenge is, or where he will pursue it.
The Economist and other sourcesShare