MADISON, Wis. — The year 2016 opened the door to a new phase of highly automated driving, moving the discussion away from “wouldn’t it be nice-to-have-a-robo-car” to a more immediate “to-do list” with which regulators, car OEMs and technology companies must grapple if they hope to make self-driving cars commercially viable and safe.
Gone are days of early-adapter giddiness over the Google car, or an “Autopilot” Tesla with over-the-air software upgrades.
Reality sank in 2016. The industry is now aware Autopilot’s limitations. The automotive engineering community is taking a crash course in Artificial Intelligence (AI) that’s far beyond today’s computer vision. Engineers are taking note of challenges in machine learning (how do you certify the safety of AI-driven cars?). Many automakers are scrambling for a holistic approach toward cybersecurity.
So, what’s in the auto industry 2017 agenda that could change the course of robotic car development?
EE Times predictions? The industry will 1) embrace open-source software for robotic cars (yes, they are coming), 2) define AI-ready SoCs (inference engines) for autonomous driving, and 3) explore a new methodology that could test and validate the safety of AI-based cars. Open-source software for autonomous cars Perhaps the most controversial issue is the emergence of the open-source software movement.
For Big Auto, whose modus operandi has been a closed, go-it-alone approach to technology development, using open-source software might be a gear too far. But advocates counter that the open-source community is much better at solving really hard problems. Autonomous driving is a really hard problem.
Phil Magney, founder and principal advisor for Vision Systems Intelligence (VSI), agrees.
Magney pointed out that “software is the golden ticket for the future of automotive.”
“In my opinion,” he said, “open source will accelerate the development of autonomous vehicle technologies. The open source community has the ability to solve tough software problems much faster than under traditional methods. Companies can then use open source code and build their solutions around it.”
But how on earth do you get started on open-source robo-car software development?
For starters, consider Comma.ai, a startup now providing the source code for OpenPilot to all comers. George Hotz, Comma.ai’s founder, who calls his new project the “open-source alternative to Autopilot,” is distributing the Comma.ai code through GitHub.
If you recall, Hotz surprised the industry last fall by cancelling the launch of his first product, an aftermarket ADAS unit, after U.S. regulators posed safety concerns.
But Hotz wasn’t deterred. He doubly shocked the industry in late November by deciding to give away the code for his self-driving car project.
Magney explained to us, “Some parts of the [Comma.ai] code cannot be modified, such as Comma’s trained network – a.k.a. inference model.” Other parts are open, like HMI and the control commands, he added. “This platform is for development only. If someone wanted to commercialize they would need to negotiate. OpenPilot is really the behavior model based on Comma.ai’s trained network,” he noted.
Good point - there is an "open-vehicle" HW website btw: http://www.openvehicles.com/vehiclesupport
The supported HW listed is Tesla Roadster, Volt and Renaul Twizy. The Twizy has only 2/5 points in the NCAP crash test. As an engineer I'd probably prefer the other two for self-driving experiments.