Here's a race condition we can get behind: Neural net learns to keep up with 'skilled' amateur track driver in robo-ride safety experiment
Watch this science friction in action
Video Researchers claim to have trained an autonomous vehicle to drive as well as an amateur race-car driver, a skill set that could be used to build safer artificially intelligent motorists, in theory.
Boffins at Stanford University in the US taught the computer brain of a driverless Volkswagen GTI to negotiate the Thunderhill Raceway Park in California at speeds of up to 95 miles per hour. Its skill level is said to be roughly comparable to that of a good amateur driver – well, one with an otherwise entirely empty race track to itself.
More practically, the researchers said their work could be used to make future self-driving cars safer, because they have essentially built a neural network that has experience with a range of road conditions, from rough asphalt to ice. This knowledge, we're told, allows the software to intuitively keep control of the vehicle in situations it hasn't experienced before. If that's the case, this competence could be built into next-gen self-driving systems to make them better at driving than their human owners.
“Our work is motivated by safety, and we want autonomous vehicles to work in many scenarios, from normal driving on high-friction asphalt to fast, low-friction driving in ice and snow,” said Nathan Spielberg, first author of a paper describing the project, which was published in Science Robotics on Wednesday.
“We want our algorithms to be as good as the best skilled drivers – and, hopefully, better.”
Car crashes are, after all, mostly down to human error, Spielberg, a graduate student in mechanical engineering at Stanford, noted. The academics reckon 94 per cent of them are the result of “human recognition, decision, or performance error.” If an autonomous car can take over in extraordinary situations, such as when a car needs to suddenly swerve, speed up, or brake, crashes could be averted by taking humans out of the loop – assuming the machine-learning software can do better than people.
Self-driving cars doomed to be bullied by pedestriansREAD MORE
So, to test this theory that algorithms can make more skillful drivers than humans, the crew trained a feed-forward neural network to whiz the Volkswagen around an oval shaped track as fast as possible without going off course. The model was taught how to stay in control of the motor from data collected during laps of Thunderhill and an icy test track near the Arctic Circle, and from 200,000 simulated scenarios using a physics engine and a wide range of virtual environments.
The aim was to teach the model to follow its chosen trajectory no matter the conditions around it: black ice, rough patches, bumps, and so on. Once the software had selected a particular trajectory to take during its lap, the network's prior knowledge was applied to stay in control. The goal was not to produce a generic all-purpose self-driving car brain, rather a primitive one that understood physics well enough to keep the car on the line needed to get it round the track.
Input data in the form of the car’s yaw rotation rate, lateral and longitudinal velocity, steering angle, and longitudinal force was fed into the neural net, which then predicted the ideal steering direction, and how fast it should travel over a specific period of time, to continue on in its trajectory as planned. This information was then used to control the car’s acceleration, braking, and steering motion.
The car's control system was trained using TensorFlow on an Intel Core i7 processor and Nvidia GeForce 1080 graphics card, we're told, with a single learning process taking 25 minutes using mini batches of 1,000 samples per update. Here’s the self-driving car, nicknamed Niki, in action using the system at Thunderhill:
After completing the course over ten trials, Niki's times compared favorably to a "skilled" human amateur racetrack driver with years of experience and knows the track well, we're told. The researchers even went as far as saying the times achieved by the neural network showed its performance is “comparable with a champion amateur race car driver.” On average, the human driver made it from start to finish in 39.9 seconds, about a second or so faster than Niki. Speed is used to measure skill, here, we note.
The team also rigged up the car to circle the track using a non-AI software controller that was pre-programmed with the course layout and details of the conditions, with the goal to drive as fast as possible around the track without crashing. This dumb autonomous system turned out to be as good as the neural network, though the machine-learning rival was driving from its experience and training rather than being hardwired for the course. In mixed surface friction conditions, or separate high and low friction conditions, in simulations, the neural network smoked the non-smart controller, we're told.
Register lecture: Teaching self-driving cars how to be more humanREAD MORE
The boffins admitted their neural network was limited to performing in the scenarios it was trained on. It is a rudimentary self-driving system: it's designed to put its foot down, and steer around race tracks using sensor data and its simple understanding of road physics to correct itself as it goes, and that's about it.
Thus, the software may be able to do well on empty courses, come snow or shine, but it won’t be able to cope with driving in traffic nor negotiate around towns and cities. Similarly, an actual race, with all that ducking and weaving with other drivers, would flummox the system. And the racing track isn't what you'd call twisty and turny: it was a loop. On the other hand, how many people do you know who could keep up with it?
Overall, it's an academic research project that at least looked like a lot of fun, the findings could help steer future robo-ride technology, and it can only improve. A Lewis Hamilton 9000 bot, it ain't... yet. Perhaps.
"We hope to design safe automated systems that are capable of fully utilizing the friction between the vehicle's tires and the road," Spielberg told The Register. "Utilizing this full capability can help autonomous vehicles of the future to avoid obstacles and collisions while staying on the road when friction is limited." ®
Sponsored: Becoming a Pragmatic Security Leader