Brazilians win int’l self-driving car challenge
A self-driving car programmed by scientists from the University of São Paulo (USP) won the international competition Car Learning to Act (Carla) autonomous driving challenge, sponsored by leading self-driving vehicle technology companies to increase their safety.
Through online platform Carla, the world’s 69 best research laboratories took part in the unprecedented competition, which tested simulated car performances. A distance of over 6.5 thousand km was traveled for more than 5.7 thousand hours.
Three of four
The USP team had the best performance in three of the four categories of the challenge. The team also came second in the only category it did not win, and received a total award of $17 thousand. “We kept our focus on lowering the number of infractions the car would commit over the course of its trip as much as we could,” said Iago Pachêco, a master’s student at the Mobile Robotics Laboratory of the Mathematics and Computer Science (ICMC).
The international competition required the 211 participants to have their vehicles go through virtual roads facing traffic, rain, traffic signs, lights, reckless drivers and pedestrians, and other unexpected situations.
The teams had to program vehicles in their laboratories and send the codes to the computers processing the data. The platform, in turn, would check how each vehicle had behaved and gave them their score accordingly. Unlike what happens in rally games, the winner was not the fastest car, but the one with the fewest violations.
Sensors
The four categories in the Carla challenge were based on the sensor types available. Some had very little equipment—just a camera and a GPS—others collected data through laser sensors, for instance. “In each category, the vehicle had to travel through a number of routes. The score was calculated according to the average on each route,” USP PhD student Júnior Rodrigues da Silva explained.
He pointed out that the vehicle’s automation structure starts with perception. “We had to design perception algorithms capable of perceiving the world in a way similar to how humans perceive it,” he said. The vehicle must know, for instance, when the lights are red or green, its distance from the traffic lights as well as the car up front, among other pieces of information.
Decision making
The next challenge is to work on decision-making codes, “to gather all the perception data to make the best decision,” Rodrigues noted. “This information allows us to execute the control layer, which will turn this command into action. The car then slows down and comes to a halt,” he explained.
The performance of cars in the competition generates statistics that helps identify the level of research on car automation. “It’s a public platform everyone can access. This allows us to compare different approaches for self-driving vehicles. Each laboratory has a different style, different algorithms, for how to work with the vehicles,” Iago Pachêco went on to say.
The USP scientists will now use the knowledge acquired in the competition on real vehicles. “We have an operational vehicle, but during the competition we developed some complements that didn’t feature on our current vehicle. The idea is to move them to the current vehicle and assess its performance in real-life situations,” Pachêco noted.
Advantages
Among the reasons to invest in research on vehicle automation is accident reduction. “Studies shows that the main reason behind an accident is a human miscalculation. In an autonomous system, vehicle intelligence can better tackle unexpected situations and decrease the source of problems,” Pachêco argued.
Other reasons include providing access for people facing driving difficulties; urban mobility, as cars make better use of roads, reducing traffic jams; the time–trip ratio; and higher energy efficiency.
