Uber’s Pittsburgh Experiment, featuring semi-autonomous vehicles, is up and running. If only its fleet could distinguish the proper path down a one-way street. And Google is reporting smashing results for its autonomous vehicle program.
This is a public service alert for all you Yinzers out there: Get off the road; you’re in danger.
While we’re at it, to unemployed tech bros desperate to get a foot in the Silicon Valley door: Don’t take a gig as a Google autonomous vehicle test driver.
There were no injuries in this incident, but one of Uber’s self-driving cars went the wrong way down a one-way Pittsburgh street on September 26. Here’s the video, posted to Facebook by Uber driver Nathan Stachelek.
Here’s more: A test driver “operating” a Google Lexus-model autonomous vehicle on September 23 was fortunate to escape with no serious damage.
Free Reports:
But he did check himself into a local hospital for evaluation after getting T-boned by another car. The Google autonomous vehicle did have the green-light right-of-way through the intersection, and the other car did run a red light. The operator assumed control of the vehicle and tried to override the decision to stop mid-intersection. A collision ensued.
Is the combination of these incidents cause to put the brakes to the runaway fervor about autonomous vehicles?
It won’t.
This is a public service alert for all you Yinzers out there: Get off the road; you’re in danger.
But events like this illustrate the complexity of artificial intelligence (AI)-human interaction in a real-world setting.
This is really hard – as in “difficult” – science. We’re talking about solving and integrating concepts such as computer vision, deep learning, machine learning, and latency.
Uber is field-testing its self-driving technology in Pittsburgh for a pretty solid reason: It’s where Carnegie Mellon University is located, and Carnegie Mellon University is a hotbed for robotics research.
It’s where the three founders of Anki went to school. Marc Andreesen, the founder of Netscape and a successful venture capitalist, named Anki “the best robotics startup I’ve ever seen.”
Anki’s breakthrough is an AI robot named Cozmo.
Robots and self-driving cars/autonomous vehicles have a lot in common, including a central problem.
As Chelsea Finn, a PhD candidate at the University of California, Berkeley, put it, “The number one challenge is to see the unstructured environment and make actions depending on the state of the environment.”
The importance of understanding and making decisions in an organic environment is perhaps best illustrated by an extreme example.
Take the military context, where using self-driving cars and trucks could save lives and reduce burdens on soldiers.
Dr. Bob Sadowski, Army Chief Roboticist at TARDEC Ground Vehicle Robotics, told Colin Clark of Breaking Defense during the 2016 Association of the U.S. Army Annual Meeting and Exposition that his chief concerns are “how you fuse the data” and teaching robots how to do this, because “they don’t know anything.”
Clark’s interview with Dr. Sadowski is here.
This is really hard – as in “difficult” – science. We’re talking about solving and integrating concepts such as computer vision, deep learning, machine learning, and latency.
The Army’s current system integrates cameras and sensors to gather data and computers to process same. Cameras “are great for giving you visual information.”
“But,” as Dr. Sadowski notes, “they’re not good at giving you 3D depth.” That problem is solved with a remote sensing method known as LIDAR, or “light detection and ranging.”
An AI sees “dot plots” created via LIDAR that represent objects in the distance. It has to learn, or be taught, to distinguish trees from humans, for example, or something else it doesn’t want to run into.
Visual information and the LIDAR information are “fused” to establish a “cost map,” which illustrates where the vehicle does or doesn’t want to go – into a tree or through a human, for example.
The potential applications are compelling, hauling equipment or even conducting reconnaissance.
The former task requires “just a little autonomy,” because soldiers will always be nearby, escorting the load, as it were. The Army is “within five to 10 years” of deploying such technology.
Reconnaissance is another matter entirely. The whole point is to spare soldiers from dangerous, life-threatening work. That’s full autonomy. And the AI would have to be able to assimilate an “unstructured environment.”
Google, Uber, Tesla, and other entities working on autonomous vehicle technology have the benefit of working in what Dr. Gadowski describes as “structured environments.”
San Francisco and downtown Pittsburgh, he notes, probably aren’t going to change too much. They can be mapped, reliably, with help from the GPS network and the continuously evolving “cloud” network that’s full of more and more usable information.
The potential applications are compelling, hauling equipment or even conducting reconnaissance.
The Army’s environment is “ill-defined,” and there are “usually people shooting back at you.”
There’s also the problem of sabotage. During World War II, the German Wehrmacht’s Operation Grief included a scheme to reverse and otherwise muck up road and traffic signs and signals to confuse advancing Allied forces.
Consider, then, the recent difficulties Uber and Google have encountered in “structured” environments with stop signs, red lights, and other mechanisms that establish some control and predictability.
Autonomous vehicles are an exciting concept. According to Lux Research, self-driving cars will represent “an $87 billion opportunity in 2030.”
At the same time, even Lux concedes “none reach full autonomy” by that date.
It boils down to AI and machine learning.
And according to Carnegie Mellon computer scientist Phillip Koopman, an automotive industry specialist, “You just can’t assume this stuff is going to work.”
Smart Investing,
David Dittman
Editorial Director, Wall Street Daily
The post Artificial Intelligence: The Race Is On to Smarten Our Cars appeared first on Wall Street Daily.