Author: John P.Desmond, AI Trends Editor
Raquel Urtasun started as an entrepreneur on June 8 with the announcement of his autonomous driving software company Waab, backed by $ 83.5 million.
Urtasun has a long experience as a researcher in computer science, and he works especially to apply artificial intelligence to self-driving car software. Uber hired him in May 2017 to lead a research team based in Toronto for the company’s self-driving program. (Look AI trends, June 29, 2018)
“Self-driving is one of the most interesting and important technologies of a generation. When resolved on a scale, it will change the world as we know it, ”said Urtasu at the launch of Waab. Press release. “Waabi is the culmination of my life’s work to bring commercially viable self-directed technology to society, and I am honored to join the ranks of additional scientists, engineers and technologists who are equally committed to realizing this bold vision.”
Waab’s launch was received with some skepticism, given the state of the self-driving automotive industry, which seeks to leave the country. But Urtasan knows what he is doing.
The latest round of funding was led by Khosla Ventures, which was also attended by Urtasun’s former employer Uber and Aurora, an AV startup that ended up buying Uber ATG last year in a deal. Limit. Money was also raised 8VC, Radical Ventures, Omers Ventures, BDC, AI Geoffrey Hinton, Fei-Fei Li, Pieter Abbeel, Sanja Fidler and others, the report reports.
Waabi will initially focus on the truck industry by offering its software to automate driving on commercial delivery routes. One reason is that there is a shortage of truck drivers in the industry. Second, highways are simpler than city streets for autonomous vehicle navigation.
Wasab’s technical approach relies heavily on simulation using techniques developed by Urtasan in his research. The company’s simulation approach reduces the need to run miles of tests on actual roads and highways recorded by autonomous competitors
“For us in the simulation, we can test the whole system,” Urtasun said Limit. “We can train the whole system to learn in the simulation, and we can produce the simulations with incredible accuracy so that we can really correlate what happens in the simulation with the real world.”
Independent vehicle starting by a woman who has developed technology is unusual; Urtasan hopes to inspire other women to join the industry. “This is a field very well dominated by white dudes,” he said. “The way to build integrative knowledge is to build technology with different perspectives, because by challenging each other, we build better things.”
Previous career at Uber, Toyota
Urtasun started in Uber in May 2017 to continue machine detection of self-driving cars. The work includes machine learning, computer vision, robotics and remote sensing. Prior to attending college, Urtasun worked at the Toyota Institute of Technology in Chicago. Uber pledged to hire dozens of researchers and pledged a multi-year, $ 1 million commitment to the Vector Institute in Toronto, which Urtasun founded.
Urtasan has argued that self-propelled vehicles must distinguish themselves from Lidar (light detection and distance), a remote sensing method that uses a pulsed laser to measure varying distances. His research has shown in some cases that vehicles can get similar 3D data from around the world from standard cameras, which are much cheaper than Lidar units that cost thousands of dollars.
“If you want to build a reliable self-propelled car right now, we should use all possible sensors,” Urtasun told Wired interview published in November 2017. “In the longer term, the question is how can we build a fleet of self-propelled cars that aren’t expensive.”
The company’s technical “artificial intelligence-first approach” means they place more emphasis on machine learning models and less complementary technologies, including Lidar, radar, and mapping data. TechTalks. “The advantage of a software-weighted stack is the very low cost of updating the technology. And there will be a lot of updates in the coming years, ”said Ben Dickson, author of the report and founder of TechTalks.
Urtasun described the artificial intelligence system used by the company as a “family of algorithms” in a 2000 report TechCrunch. Its closed-loop simulation environment replaces sending real cars on real roads.
“I’m a bit of a simulation component on the fence,” Dickson noted. “Most self-driving car companies use simulations as part of their in-depth learning model training system. But creating simulation environments that are real imitations of the real world is virtually impossible, which is why self-driving car companies still use a heavy road test.
Waymo leads in simulated and real test miles
Waymo has at least 20 billion miles of simulated driving with 20 million miles of actual road testing, which is an industry record, according to Dickson. To learn more about Waab’s technology, he looked at some of Urtasun’s recent academic work at the University of Toronto. His name appears in many papers on autonomous driving; one uploaded to the arXiv preprint server in january caught Dickson’s attention.
Title “MP3: A unified model for mapping, detecting, predicting and planning” the article discusses an approach to self-driving close to the description in Waab’s release bulletin.
The researchers describe the MP3 file as “an end-to-end approach to mapless driving that is interpretable, does not cause data loss and causes uncertainty in intermediates.” In the study, researchers also look at “the use of probability-based space layers to model static and dynamic parts of the environment.”
MP3 is end-to-end trainable. It uses Lidar input to create scene representations, predict future states, and plan routes. “The machine learning model eliminates the need for detailed map data that companies like Waymon use in their self-propelled vehicles,” Dickson noted.
Urtasun posted a video, The future with self-propelled vehicles, on his YouTube channel, which provides a brief account of how MP3 works. Some researchers comment that it is a smart combination of existing techniques. “There is also a significant gap between academic artificial intelligence research and applied artificial intelligence,” Dickson noted. How the Waabi model works in practical conditions is interesting to watch.