Mitsubishi plans to include scene-aware interaction in cars
spectrum.ieee.org
spectrum.ieee.org
Sounds better than what we have now, which is dumb systems that often repeat stuff unnecessarily and annoyingly: “In a quarter mile, turn right” “In 500 feet, turn right” “Turn right”
I want it to pick one and shut up unless I’m about to miss the turn: “Turn right at the next light” “Are you going to turn or what?”
This would be less annoying and also reduce the amount of assistance my wife feels compelled to provide.
Ed: s/nab/nav/g
Easy solution here would be to have some feedback mechanism for the nav. Surprised it doesn't exist already. I miss what the nav says about half the time, because I'm paying attention to other things. But if i could say "ok, got it" or press a button that says "ok, got it" the first time she gives a new direction, that would shut up her reminders, that seems ideal.
Plus just turn the "turn in 200ft" to "turn in two blocks". None of this needs fancy sensors.
Similarly, I'd like to be able to say something like "get me to I-71 North" or similar and have it take me to the nearest exit for that direction.
It's the main reason I won't use Google Maps any longer.
It's harder to reason about 0.2 mi, 700 ft, 400 ft, 250 ft, esp when streets are close together, or at different approach speeds.
yes. the apple navigation app can stop saying "proceed to the route" after 40 or 50 consecutive utterances - 100 is just excessive. It can get annoying during a road trip when you want to randomly pull off to take a look at something.
Maybe if there were just a voice command: “Hey, Widget, mute nav for 15 minutes.”
Apple maps says things like "at the next traffic light, turn right onto Crankyoldman Street." It's been a while since I used Google Maps.
You can just turn off prompts entirely. Between CarPlay / Android Auto or just having your phone in a cradle relatively high-up on your dash, voice prompts aren't necessary if they're annoying you that much.
I actually find these super helpful. They're redundant in some cases, but if you ever fail to hear what they say the first time for whatever reason, or if there are multiple streets close to each other, the redundancy is incredibly helpful.
No more "500 feet". Instead it says things like:
"Go through this light, and at the next one, turn right."
Not sure if its routes are always the best but I tend to pick it anyway because of the more intelligible instructions.
Then there's my perennial favorite: "You... will... reach... your... destination... in... three... hundred... feet. ... Your... destination... is... ... was... on... the... right.
There is so much low-hanging fruit in the automotive nav business, so much stupidity, so much room for improvement that never seems to occur to anyone to implement. It's just incredible.
- so that trope will live on!
Because the vehicle’s awareness of its surroundings is transparent to passengers, they are able to interpret and understand the actions being taken by the autonomous vehicle. Such understanding has been shown to establish a greater level of trust and perceived safety. "
I mean - no? Google can’t track more than one timer at a time (and can’t subtract time from it either) and recognizes me maybe 80% of the time and my wife maybe 15%. Alexa let’s strangers talk to your kids and replies in memes and farts, Siri genuinely barely understands English, and I haven’t used Microsoft’s video game character but at least the other three are so far from “impressive” it’s hard to take the article seriously with an intro like that.
Is anyone impressed by the state of voice recognition or the utility of voice assistance?
I was hoping for some new take on the AI/ML blackbox approach that could allow for us to trace a vehicle decision back to deterministic rules. Ya know, something like SQL.
This is still interesting on the UX side of the house though. A lot of people navigate better this way. It might actually reduce cognitive load and accidents by providing information in a more digestible format.
So, this is for the subpopulation of car drivers that have no accurate sense of distance? Mmh, okay.
> - follow the silver car to turn left
> - I can't see it
Not surprising. Colors are somewhat ambiguous, cultural and subjective. Greyish ("silver") is a very common color for a car. At this moment there were several car matching that description.
So, using color, at least in this case, is not accurate.
At least, with a distance, the information is somewhat well-defined.
> - okay follow the silver car in the leftmost lane
That's another silver car.
Wait, what's the benefit of mentioning that car, compared to saying "join the leftmost lane and follow it to turn left"? Or simply, "turn left".
Do they mean it is for drivers that can follow another car but not painted lanes?
What happens in low traffic when there's no other car to follow?
> turn left at the building with a billboard
There are several buildings with a billboard. One at around 100 meters, one at around 300 meters.
So, mentioning "a building with a billboard", at least in this case, is not accurate.
At least, with a distance, the information is somewhat well-defined.
> recent breakthroughs in object detection and recognition semantic segmentation motion analysis of dynamic objects and natural language processing technologies our system also makes use of high precision maps
"recent buzzword in buzzword and buzzword buzzword buzzword buzzword of buzzword and buzzword buzzword our system also makes use of buzzword".
> surveillance systems that understand complex scenes and interpret them for humans
Ah, this may be the real deal. No, thanks.
I also happen to sometimes miss guidance directions (happened a few days ago). Still, I doubt ambiguously referring to environmental features is an actual progress.
My feeling is that in both current and projected systems, the driver will need some experience to figure out the ambiguity.
I also anticipate cases where the AI will mislabel an object and ask to turn left at the X while there's clearly no X, or maybe actually an X but further away, causing misguidance.
Not that I think you need to apologize, but if-pologies are worse than nothing. They irritate people more.
> Why do the media keep running stories saying suits are back? Because PR firms tell them to.
> Communication between humans and machines is hindered by a lack of shared awareness.
Yeah... by the fact that humans aren't aware that AI is just a bunch of 1s and 0s and not actually aware. There's no intelligence here.
> Scene understanding
... it cannot understand. It can only match. It does not understand in the way humans understand.
And the human brain is just a bundle of electrical signals. Why's that any different?
Consciousness, sentience and sapience. Which is to say, that is indeed the multitrillion dollar question.
What makes something conscious?
Follow up: is that answer well-defined (it's not using other nebulous concepts to avoid having to rigorously define it), and does it accept all of the things you think should be called "conscious" and reject all of the things you think should not be?
the more we look the more we find that this is not the case. and i don’t have to invoke any philosophy here, just remind you that neurotransmitter modulation and glial cells exist.
Honestly, the root of all this is whether you were trying to claim that the brain is doing super-Turing computation or not. Either
1) it's not, then you can't tell me Turing machines cannot do the same computation (and by extension neural nets can approximate it)
2) it is, then prove it and congratulations you have made a extraordinarily important and fundamental step forward in computer science.
My money is on 1.
Certainly neither of them are action potentials, which are usually the (exclusive and therefore imo incomplete) basis for the mental model of people who claim things like "the brain is just electrical signals." No. It's not. It's a vastly more complicated system of organ(s) than that even if you're willing to draw a line in the sand between brain and environment -- and that's no easy feat itself!
This is relevant because the universal function approximaters we are training everyday are capable of representing this computational model, undermining the argument that there is somehow a difference between the model of computation done in the brain and the computation done by neural networks.
oh. well, that’s just a naive embrace of the andy grove fallacy.
as of this moment, the brain’s computational model (and if there is one, and if so how many) is neither well understood nor replicable.
it’s almost a category error to describe the process as “reverse engineering”, but even if that’s an appropriate mental model, it turns out that reverse engineering an evolved system that’s taken 4.5 billion years and has had no human design additions at all has very little in common with any similar approach you could take in, say, inspecting the trained weights of a neural network to back out what’s going on.
No it's not. That's about feasibility, and the comment was not about feasibility. It was about turing vs. super-turing.
Go ahead and assume the reverse engineers get 4.5 trillion years or something. That wasn't the point of the "actually aware" rebuttal.
And?
If it can match to a point where it knows exactly what everything is, and presumably is capable of (or eventually will be able to) making decisions based on that information well enough to obey the rules of the road, then it doesn't really matter if it's using trillions of simple logic gates or a human brain.
Human understanding enables inference. Say your full self driving car is driving down the freeway and an enormous explosion / forest fire / volcano erupts in the distance, it'll blissfully drive towards the disaster - because it has no comprehension of what the orange sky, shaking ground, fleeing people imply.
This is a extreme example but the more we learn about driving the more we realise our ability to do it is based on inference not only about the world, but about the intentions of other drivers.
There's no 'decision' being made by a self driving car - it's matching its predefined set of visual inputs to a set of behaviours. It can never scale to adapt to the unknown - even in principle.
> It can never scale to adapt to the unknown - even in principle.
But the human will be able to fill in for the inference cars don't have. Even if we get to level 5 where there's no steering wheel, the human can still control the car's destination and what sort of operation modes the car is in; if there is some event and full-scale panic ensues, the AI driver isn't going to hit people in the street - either the human can instruct it to turn around, or they can get out of the car, since even now it's not really safe try to weave between humans in the street to reach your destination.