Computer science professor Raquel Urtasun, Canada Research Chair in Machine Learning and Computer Vision, is creating the blueprint for an achievable, affordable self-driving car.
When you trace the history of neural network research, all roads lead back to one professor and one university: Geoffrey Hinton and the University of Toronto. Hinton, often called the godfather of deep learning, has led groundbreaking research in machine learning—teaching computers how to think—for over 40 years. Hinton’s work has been widely acclaimed and adopted by the tech world, including Google, which recently acquired his start-up company DNNResearch. As we enter an era where artificial intelligence exists in everyday life, a new pioneer has come to the fore at the University of Toronto. Professor Raquel Urtasun is drawing a new map, and thanks to her work, our roads will soon be safer.
After completing postdoctoral fellowships at MIT and UC Berkeley, Urtasun was interested in pursing a number of different areas: machine learning, computer vision (enabling computers to understand the world through imagery) and robotics. She decided to focus on all three by developing the algorithms and conditions required to create safe, sustainable self-driving cars.
For many years, the prospect of the self-driving car was proving unrealistic and unaffordable. To date, self-driving cars have relied on expensive sensors such as LIDAR, which puts their cost at over $100,000—unrealistic for public use. These models also require manually annotated maps, which are incredibly expensive to produce. Urtasun explained, “Experts have calculated that to map the whole United States would cost more than one billion dollars. Imagine the cost of keeping the maps up to date.”
Now, Urtasun’s research is paving the way for an achievable approach to self-driving cars. Affordability is key, as her vision is for autonomous vehicles that could be used everywhere, helping make our cities safer.
“Over one million people die every year due to automotive accidents,” she said. “Imagine the impact if we could reduce that number significantly.” With Urtasun’s advanced software, self-driving cars would understand and adapt to their environment and surrounding vehicles, drive less aggressively, and ultimately lower collision rates.
At a special TEDxUofT event, Urtasun spoke of the potential environmental benefits of self-driving cars. “Self-driving cars could help to reduce pollution. They could present new opportunities for faster, safer public transport and a more creative use of space.”
Urtasun looks to self-driving vehicles as a way to encourage more people to use public transport, car sharing and ride sharing.
“Owning a car is really a waste,” she said. “We should be thinking of sharing resources, by way of materials and also space. Ninety-five percent of the time a car is parked and unused, it is such a waste of space. Imagine what we could use all of that space for. We could use the space that cars take up to make our cities much better.”
One of the problems with public transport is its lack of flexibility. Urtasun believes that self-driving cars could allow public transport to respond to demand. “They would make commutes faster and easier for people who live outside the main arteries of a transport system,” she said. “You could maximize the number of people per car—which is the big problem currently, as we have all these commuters who travel in cars alone.”
Urtasun’s research has focussed on using inexpensive sensors to interpret information captured from a camera. “As humans, we can solve this problem. We use our eyes when driving. In the same way, a computer could use a camera.” For computers to interpret and understand the world as it’s presented through a camera, there are many obstacles to overcome. Computers will need to understand 3D and the proximity and movement of surrounding vehicles.
Urtasun spent the first two years of her research on self-driving cars building the benchmarks that will enable developers to create the algorithms required to create autonomous vehicles. She and her team developed the KITTI benchmark, a series of challenges for developers that are used to simulate and assess the performance of their systems. The benchmark has been used by over 500 groups in both academia and industry, including Daimler, NVIDIA, Samsung and Toyota.
To date, Google, NVIDIA, Intel, Samsung, Toyota, Mitsubishi, Adobe, the HSA Foundation and Amazon have all supported Urtasun’s research. She was also recently awarded the E.W.R. Steacie Memorial Fellowship by the Natural Sciences and Engineering Research Council of Canada. The fellowship is awarded annually to enhance the career development of outstanding and highly promising scientists and engineers at Canadian universities. In addition to her role at the University of Toronto, Urtasun was recently recruited by Uber to lead its self-driving research in Canada.
Toronto is fast becoming a global hot spot for machine learning, with the launch of a new independent centre for artificial intelligence in Toronto. The Vector Institute, which is affiliated with the University of Toronto, advances Canada’s claim to be the global leader in artificial intelligence. Urtasun is a co-founder of the institute and will be a key member of its research team. And Geoffrey Hinton will serve as its chief scientific adviser.
As interest in her work continues to grow, Urtasun’s team moves closer to making self-driving cars a reality. But for her, this is just the beginning. “Self-driving cars aren’t the end goal. This is the first step towards having really smart cities.”
Story by Phil Boughton
(top) Raquel Urtasun. Photo: Jacklyn Atlas.