Friday, September 14, 2007

2007 Autonomous driving systems aim to drive dirty

Autonomous driving systems aim to drive dirty

Movie Camera
  • 17:36 23 March 2007
  • NewScientist.com news service
  • Matthew

Autonomous model cars will race against one another in a contest designed to test different software approaches.

The contest is being organised by researchers at the University of Essex in the UK, who are creating an affordable and standardised autonomous vehicle kit to encourage others to get involved.

The kit will include a high-end commercial model car, a laptop, a GPS receiver, a USB controller and a camera. The aim is to encourage different research teams to develop autonomous racers using the same equipment, which will then race against one another at the 2008 World Congress on Computational Intelligence in Hong Kong.

Simon Lucas of Essex University says the competition will be similar to the DARPA Grand Challenge (see Desert racers – drivers not included), which involves full-sized vehicles, but will be far less prohibitive. "The challenges are the same for a full-size or model autonomous car, but you need pots of money," Lucus told New Scientist. "Our prototype hardware costs only £1000 ($2000)."

The contest will be preceded by a simulated competition held at the 2007 IEEE Symposium on Computational Intelligence and Games in April 2007. In this competition, the software used to control a car will be run on a simulator. The goal is to reach as many waypoints along the track as possible in a given time without crashing into a second virtual vehicle.

Aggressive racing

Competitors will be able to apply different programming techniques to autonomously direct their cars. Lucas and PhD student Julian Togelius plan to use software built around a learning, "evolving" algorithm.

Developing this involves testing hundreds of different algorithms against one another, then selecting the best ones for recombination and mutation – a process that mimics biological evolution. Videos created by the researchers shows the virtual cars in action.

The software developed by Lucas and Togelius has already demonstrated an ability to complete a course faster than a human controller and has even learned to drive aggressively, knocking into other simulated cars to achieve an advantage.

Lucas says a race involving actual cars, albeit model ones, will require greater skill. "The challenge is to use computer vision methods together with a range of other sensor data to race the car as fast as possible around the track while outwitting the opponent cars," he says. "To do so it needs to be smart and it needs adapt to the behaviour of the other cars as it drives."

No comments: