By Michael Lachmann
Producer, BBC Horizon
Twenty-three cars are setting off on a remarkable race across the desert.
"Stanley" with Stanford University team member Sebastian Thrun
From a distance, the cars look like any other rally cars tearing down the dusty roads. But closer inspection reveals that something odd is going on.
There's no-one sitting in the driving seat of any of the cars. Nor are they being steered by remote control.
This is the Grand Challenge, the world's toughest race for robot vehicles - cars that drive themselves. There is $2m (£1m) on offer to the car that completes the 135-mile (220km) course in the fastest time.
To win it, the cars will have to drive 20 times further than any robot car has ever driven itself before.
The race has been organised by Darpa, the US government's military research arm, who are keen to develop driverless technology.
"If we can use unmanned vehicles in dangerous missions like delivering supplies," says Ron Kurjanowicz of Darpa, "then we can take the drivers out of those vehicles and save lives on the battlefield."
For Sebastian Thrun of Stanford University, who is the designer of one of the favourites, the technology has wider appeal. "The competition," he says, "is to make the world's roads safer.
"The day when my car commutes for me and I can sit there, read the newspaper, do e-mail in the car while the car is driving itself, that's the ultimate victory."
The winning team celebrates its victory
Over the opening miles of the course, Thrun's car, Stanley, makes that ultimate victory seem close at hand, sailing smoothly down the unpaved roads at up to 50mph (80 km/h).
Built on the body of a VW Touareg, it has been fitted with a drive-by-wire system that allows the car to be operated by computer.
The steering, gear changes, throttle and brake are all controlled by the robot itself. But driving the car is the easy bit, the hard part is deciding where to go.
As computer engineer Jay Gowdy puts it: "No-one knows how to make a robot understand its environment, or how you make a robot see. Instead, we cheat and build things that measure."
The robots use an array of measuring devices to build a 3D map of the road ahead. Radar systems scan the terrain, while lasers identify and pin-point obstacles, and cameras look further ahead to give warning of what is coming from the distance.
Equipped with the map, the robot then just has to pick a route through it. The flattest bit of the map is likely to be the road, so the robots drive down the flattest course they can while avoiding any obstacles in their way.
It might sound easy but it takes a huge amount of computing power. Stanley uses a million lines of specially written computing code and the processing power of seven on-board computers to plot a course and decide its speed.
The difficulty of the task was graphically demonstrated when the Grand Challenge was run for the first time in 2004. Of the 15 vehicles that started the race, half crashed or ground to a halt within the first mile.
The most successful vehicle, a converted Hummer called Sandstorm, managed only seven miles - over 100 miles short of the finishing line. It was a chastening experience for its creator Red Whittaker who has returned for another go at the course.
He has no doubt how difficult the Grand Challenge is. "I think of it as robotics' Kittyhawk," he says, referring to the Wright brothers and the development of flight. "But it calls for a 100-fold leap in the performance of robotic technology."
To double their chances, Whittaker and his team from Carnegie Mellon University in Pittsburgh have entered two vehicles into this year's race, both rugged pieces of heavy engineering designed to take on the gruelling desert course.
It's a strategy that appears to be paying off as vehicle after vehicle drops out of the race. Some suffer from mechanical failure caused by the punishing terrain, but most fail as their navigation systems misidentify the road, sending the robots off course and into the desert.
Before long, the two Carnegie Mellon robots and Stanley are on their own at the front of the race. For nearly 100 miles, they contest the lead, each steering their own route and adjusting their speed as the terrain changes.
After nearly seven hours of racing the teams gather at the finishing line to see which of their robots will return first. A cloud of dust on the horizon becomes a blue dot and, to jubilant celebrations from the Stanford team, Stanley drives itself across the finish line and into the winning circle. Within 20 minutes, both the Carnegie Mellon robots also finish.
That three vehicles completed the course in such quick time astounded many observers and has bought closer the day when we are all driven around by our cars.
But there are still many problems to be solved. For instance, the robots in the Grand Challenge only had to avoid static obstacles. What will happen when the cars have to contend with moving hazards such as other cars and pedestrians?
Hopefully, we'll find out next year. The teams are preparing for a robot race called the Urban Challenge, in which the cars will have to negotiate a 60-mile (100km) course along built-up roads, avoiding other cars - and keeping to the traffic laws.
You can watch The Great Robot Race on Horizon, BBC Two, at 2100GMT on Tuesday 31 October.