“WHATEVER you do, don’t stand in front of the vehicle,” shouts Chris Urmson. He is about to test-run Sandstorm, an adapted Humvee that is the favourite to win the title of fastest autonomous vehicle in the world – not to mention a $1 million prize – in a desert race from LA to Las Vegas.
Urmson, a graduate student from Carnegie Mellon University in Pittsburgh, Pennsylvania, jumps into a separate car and pushes a button on a wireless box. Sandstorm is off. We chase its plume of red dust at around 40 kilometres per hour over rugged desert terrain, while Urmson’s finger hovers over the emergency stop button.
This weekend’s Grand Challenge race is organised by the Pentagon’s Defense Advanced Research Projects Agency (DARPA), and is the most high-profile clash the robotics world has ever seen. As żěè¶ĚĘÓƵ went to press, 20 select teams were preparing to send their vehicles off on a 320-kilometre race. The robots will be allowed no wireless help between start and finish, and the route kept secret until just 2 hours before the race. This will be a formidable challenge, and many experts believe none of the robots will finish. The race has also been mired in controversy, with mudslinging between competitors and criticism levelled at DARPA for changing the rules of engagement since the race was first announced. DARPA says the changes are little more than tweaks.
Advertisement
Its motivation for organising the race is a congressional mandate that says the army must employ autonomous vehicles by 2015. The idea is that a military model based on a Grand Challenge-worthy vehicle could be used to transport supplies to troops stationed in dangerous areas, without risking human life, or as reconnaissance vehicles sent into areas deemed too dangerous for humans. “These systems would do the dull, dirty and dangerous,” says Daryl Davidson, president of the Association for Unmanned Vehicle Systems International. “Why risk a life if you can risk a machine?”
But to do this, the Department of Defense desperately needs new ideas. The best it has done so far – at least as far as unclassified technology goes – is develop a truck that can travel off-road at a maximum of 20 kilometres an hour, but it never travelled more than 10 kilometres without human intervention. Jose Negron, DARPA’s programme manager for the Grand Challenge, says he hopes the race will find a vehicle that will do better, and give the government a chance to tap new sources of innovation. “The challenge is way past state of the art,” says Alex Gutierrez, a member of the CMU team. “This is a really hard thing to do.”
The teams selected by DARPA for this weekend’s race fall roughly into two groups. There are large, well-funded outfits hailing from robot powerhouses like CMU and Caltech. And there are amateur teams that scraped through the selection process on a shoestring. The opposing approaches have led to very different kinds of innovation.
While Urmson and a busy cadre of CMU engineers and programmers established their Sandstorm camp out in the Nevada desert in February, a cheaper, simpler solution was unfolding in an icy garage more than 2000 kilometres to the north-east. Team Phantasm, a couple of hobbyists in St Louis, Missouri, were still fitting together their entry using hardware costing a mere $30,000.
“What I can’t afford to buy, I make,” says Warren Williams. He and his teammate Bill Zimmerly gained their robot experience competing in the Battlebots contests, known as Robot Wars in the UK, in which robot enthusiasts pit their home-made machines against each other in a fighting arena. Williams is a veteran of seven Battlebots contests, and has won twice. His bots are usually fully remote-controlled, but he recently decided he could gain an edge by including a level of autonomy. Then he heard about the Grand Challenge.
Two light bulbs swinging from the ceiling of William’s garage illuminate the Long-range Autonomous Directional Intuitive Boundary-sensing Unmanned Ground, or LADIBUG. The 350-kilogram bug boasts a motor culled from an electric wheelchair, a solitary onboard computer, and ball bearings snitched from his children’s rollerblades. It could not be more different to CMU’s million-dollar, multi-computered, 3-tonne monster.
Ivar Schoenmeyr, part of the small California-based Cyberrider team, says low-budget teams have to be innovative. “The big companies were less likely to donate to smaller teams, so they had to start with a clean sheet and design from scratch. The advantage is that you can custom-make something that is specific and unique to what you are trying to do.”
One example of this is LADIBUG’s robotic recovery arms, which were inspired by the team’s Battlebots experience. Put together from an air compressor and scrap metal, they extend from its sides and should right the vehicle if it falls over. The Sandstorm team, on the other hand, has eschewed such mechanical safeguards, but has invested in a software monitoring and recovery system that will sense when an application has gone down and reboot.
Money talks
Schoenmeyr agrees that Sandstorm will probably win, but questions how innovative the vehicle is. “They have taken off-the-shelf stuff, clobbered it together and plugged it into a bank of computers,” he says, “I don’t think it is a particularly elegant solution.” Naturally, the Sandstorm team disagrees. Although much of its technology, such as its state-of-art terrain-mapping camera with stereo vision, was donated by big sponsors – often defence contractors – the team insists that Sandstorm is innovative.
For a start, says Jay Gowdy of CMU, the Humvee’s path-finding system is the first of its kind on a fast-moving, all-terrain vehicle. Sensors feed radar and laser readings to a computer program that constantly calculates the distance and height of any bumps or obstacles on the landscape. Then another on-board computer builds a very simple map that divides the terrain into squares representing “possibly traversable” and “definitely untraversable”, depending on its height. This binary map acts as peripheral vision, enabling Sandstorm to rule out certain areas of terrain very quickly and focus its more precise vision on the possibly traversable areas. It also frees up the stereocamera to do the more computationally time-consuming job of precision-mapping the areas that are possibly traversable.
Gowdy says that alternating between the two types of map, calculating which will best suit the challenge, and knitting together the hardware and software all represent useful innovation. “For most of what we are doing, the underlying tech is not new, but we are putting things together in a new way,” he says.
Without expensive lasers and stereocameras, the LADIBUG cannot build detailed maps to help plan its route. Instead it simply sets off and then relies on sonar, radar and infrared to tell it whether to change tack. “We don’t think you have to paint the environment to operate in that environment,” says Williams.
Sonar is most often used under water, where it works excellently, but this is the first time it has been used on an autonomous overland vehicle. Williams borrowed a dirt-cheap sonar distance-finder from an autofocus camera. Unlike radar, which travels in a concentrated beam, ultrasonic sonar signals spread out in air to form a large “bubble”, giving it a typical range of only about 10 metres. This is not enough for a vehicle travelling at around 60 kilometres per hour, but Williams found a way to extend the range. He dug up a paper by scientists at the University of Alberta in Canada about a headset designed to warn blind people of obstacles. The headset filters off noise and amplifies signals that look like obstacles, and so Williams decided to copy it. “We take the signals that the sonar normally disregards and we have doubled the range,” he says.
Gowdy is quick to point out the limitations of a “blind” vehicle. The LADIBUG’s sonar only points straight ahead, so if the vehicle turns suddenly, it may run smack into something before it has sensed it. “If you build a robot with no peripheral vision, it does things that look unbelievably stupid,” he says.
However, the blind bots may be helped by a controversial change that DARPA has made to the race rules. It centres on sets of “waypoints” marked by GPS coordinates. These are the markers that a racing robot must pass through in order to be deemed to have followed the race route. DARPA is planning to release these just two hours before the race begins, giving teams only the briefest chance to compare the route with 3D satellite maps of the region.
Baby steps
DARPA had said at the first Grand Challenge Conference in Los Angeles in February 2003 that the race would be 480 kilometres long and have around 1000 waypoints – meaning one roughly every 500 metres. But in January this year, teams were emailed a sample route with waypoints every 7 metres. DARPA’s Negron agrees that DARPA may have “tweaked” the rules, but says more waypoints do not necessarily make the race easier, because a robot may get confused over how to follow them all in sequence.
Participants disagree. They say following such closely spaced points makes the race more like a remote-control game than a test of autonomy. “It is very clear that a substantial amount of the race will be guided in baby steps as opposed to being a solution that is devised en route,” says Red Whittaker, leader of the Sandstorm team and veteran of CMU’s Robotics Institute. The changes may favour simpler vehicles like the LADIBUG. “The major worry I have for us is the more complexity, the more things break,” says Gowdy.
One other apparent change in the race format is the proportion of off-road terrain. At the conference last year, DARPA said it would account for around 20 per cent of the route. But Schoenmeyr says his analysis of maps released by the US Bureau of Land Management, which will be used to help clear possible routes and warn people living along them to watch out for racing robots, suggests that as little as 5 per cent of the route will be off-road. It now looks like the final choice will largely avoid dirt roads with ill-defined markings and potholes – the most difficult type of obstacle for a robot to sense – in favour of easy-to-follow main roads.
Yet it is the off-road component that has produced the most innovation. Both the LADIBUG and Sandstorm teams have had to custom-build their sensing equipment to make sure that the instruments maintain their orientation, no matter how rocky the ride. Otherwise sensors could find themselves looking at the sky, sending an all-clear instead of reporting a serious obstacle. “You have to design a vehicle that can take a really good beating,” says Zimmerly.
LADIBUG’s answer is inspired by the way remote-controlled aircraft balance themselves. Williams mounted the radar and sonar in a contraption called a gimbal – a set of nested rings mounted inside one another to allow rotation in any direction. Because the ground emits more infrared than the air, infrared sensors on the gimbal can work out which way is up, allowing it to readjust the radar and sonar whenever they tip relative to ground. To keep within budget constraints, Williams built the gimbal from plumbers’ plastic tubing and electric motors for just $500. “When CMU find out how much I spent on this, they’ll probably jump off a cliff,” he smiles.
Sandstorm’s solution to the same problem was to use accelerometers and gyroscopes. It cost tens of thousands of dollars, but it will almost certainly keep the vehicle’s brain and eyes safer and more precisely aligned. The only other gyro-stabilised platforms as precise as this were built by the military to carry guns, says Gowdy, and this is the first time the idea has been used to stabilise and position robot “eyes”. “The point is the task has never been done before, so there is a lot of innovation bringing it all together and picking out the parts,” says Gowdy.
The team’s roboticists say they learned from past projects that a gimbal alone cannot steady the sensors because its motors cannot work fast enough to counter the jolting of an off-road vehicle. Something needs to slow down the shaking. So the team housed their entire suite of sensors and eight computers in a box that floats above the main chassis on shock-absorbing springs. Dumbed-down or not, DARPA’s race has already produced some new technologies, says Gowdy.
Perhaps the greatest innovation to come out of the Grand Challenge race, however, is the idea of the race itself. When it was first announced, Negron told the assembled crowd of roboticists: “We are trying to attract the independent guys, the small-business people.” But DARPA’s tough selection process left a swathe of disgruntled roboticists who have decided to create an alternative Open Challenge, scheduled for September.
Meanwhile Williams is proof that even if the Grand Challenge has been watered down a little, the fighting spirit of the competitors is as ferocious as ever. He might even have preferred a few extra “tweaks” to the rules. “If only we could have weapons on board like in the Battlebots, there’s all kinds of things I could have done to tear them up.”
- Watch www.newscientist.com for the results
