SoylentNews is people
SoylentNews
Submitted via IRC for chromas
Toyota, Carmera will create HD maps for self-driving cars using autobrake cameras
High-definition mapping is but one part of the very large equation being worked through as the auto industry prepares for autonomy. But creating those kinds of maps for every road in the world, as one might expect, takes a lot of work. Thankfully, Toyota and Carmera think they have a solution.
Toyota Research Institute and Carmera, a company focused on HD mapping, announced on Thursday that the two are developing a proof of concept of TRI's Automated Mapping Platform, which takes data from participating cars and turns that into HD maps that could be used to help autonomous vehicles navigate.
[...] The program will start in downtown Tokyo. TRI notes in its press release that HD mapping has covered less than 1 percent of worldwide roads to date, most of which are highways. Urban and local roads are likely to see large shares of AVs in the future, so both companies believe this can speed up the whole mapping process. Of course, even if the proof of concept proves viable, it would take a lot of effort to get automakers on board with the idea -- after all, even if the data is anonymized, drivers and OEMs alike may not enjoy the idea of a camera beaming what's in front of your car to some random server farm across the globe.
(Score: 4, Insightful) by inertnet on Monday March 04 2019, @03:19PM (5 children)
I think that it would be better to develop AI that can deal with the world without the need for highly detailed maps. Self driving vehicles need to be able to deal with the unexpected, instead of a world where every detail is regulated.
(Score: 0) by Anonymous Coward on Monday March 04 2019, @03:31PM
They may not like the maps made from my driving...in the snow (when there are no other cars around) I slide all over for fun and car-control practice. Come to think of it, the cameras are probably covered with snow & ice in those conditions anyway, so maybe it doesn't matter?
(Score: 5, Insightful) by AthanasiusKircher on Monday March 04 2019, @04:14PM (2 children)
Agreed, though I'm less confident than many that this sort of AI is going to be perfected soon.
The danger with many machine learning algorithms used as a basis for AI these days is that it's hard to figure out what "rules" the algorithm is actually making decisions on. You have enormous sets of weighted numbers in some mutilayered neural network, but how do you know how that's going to react to a particular novel situation?
It reminds me of a story about ALVINN, an autonomous car first designed in the late 1980s and tested in the early 1990s. It incorporated neural network models just as a lot of AI does today.
It ran fine after being trained on lots of roads. Then one day the car went across a bridge and started going nuts, forcing the human inside to have to swerve and jerk it back on track to avoid getting killed.
Back then the much smaller amount off stored data in the AI model allowed analysis, and after some weeks they figured out what happened -- it turned out that through training, the car had learned to navigate by sensing where the grass was on the side of the road. It had been trained on road consistently with grass on the side, so when it encountered a bridge, it had no idea how to navigate.
That particular case seems like something that could be worked out today through better training. The problem with AI today though is that it's often difficult to know exactly how it will react to a novel situation, because it's difficult to know exactly how it's "understanding" the situation without a great deal of analysis.
Bottom line is that I think having detailed maps is perhaps one additional data point to give the AI so in case it is getting unusual input, it has some other data potentially to guide it. Of course, the age of the data must be taken into account, but short of things like major earthquakes or unexpected demolitions, the chance that a road is going to completely disappear or move significantly from old map data is likely small.
(Score: 2) by HiThere on Monday March 04 2019, @05:06PM (1 child)
OTOH, it's often quite difficult to figure out what "rules" the non-automated drivers you encounter are using. So the real determinant needs to be quality of results. But trying to get unfudged numbers for that is ... difficult.
Javascript is what you use to allow unknown third parties to run software you have no idea about on your computer.
(Score: 0) by Anonymous Coward on Monday March 04 2019, @06:26PM
So you advocate driving tests and licenses for automated cars? The precise "rules" aren't so much an issue as the resulting ability to drive safely.
(Score: 2) by darkfeline on Wednesday March 06 2019, @06:55AM
Our world is filled with things that should be able to deal with the unexpected, but very few things that can deal with the unexpected.
Take humans, for instance. Humans generally cannot deal with anything unexpected, they suffer acute stress and mentally shut down. Physical infrastructure cannot withstand forces they were not designed for (harmonic stresses causing catastrophic failure, unexpected earthquakes or floods, previously unknown phenomena, etc.). Having depraved sex on a chair not designed for such will break it and cause physical injury and embarrassment.
The history of human progress is a history of understanding and regulating the world, developing systems and models, building infrastructure and fixing "bugs" as they occur. Your stance is idealistic and naive and not at all practical. "In theory, theory and practice are the same. In practice they are not." There will always be unexpected situations that you cannot deal with; if you ever think you handled all the unexpected situations, congratulations, that's because you didn't expect the unexpected situations. No one expects the Spanish Inquisition.
Join the SDF Public Access UNIX System today!