The alternative would be to narrow roads and lower speed limits as traffic intensifies, mitigating the need for stop lights. Unfortunately, design speed is the priority: we do whatever it takes to make sure that cars, when they are not stopped by lights or other traffic, can safely(ish) travel at the design speed.
Go right and then go around the roundabout.
I'm sitting at an intersection trying to go left, but with the intention to turn right and 180 around the roundabout; behind someone who also wants to go left but would rather wait for an opening in both directions of traffic.
People are idiotic; traffic systems should, ideally, account for and mitigate this idiocy as much as possible. That's why streets with a impassible divider in the middle exist.
On the walk/bike angle: my vague recollection is that the Not Just Bikes Youtube channel has some recordings of roundabouts that work pretty well for non-drivers, but I can't remember specifics. Maybe just heavy prioritization of those users over drivers.
It's a pretty bike friendly town though, so people here watch out. I probably wouldn't try that in a more urban city.
Additionally, you're dealing with the effects that I outlined earlier in the paragraph, but at every single intersection; oftentimes, and especially at intersections where one direction of travel (say North-South) is significantly more popular than the other (e.g. West-East), traffic basically never stops because there's little crossing-traffic to stop drivers from entering the intersection for even a second; and drivers are not required to yield to pedestrians. You start forming prejudices against car brands real quick walking around cities like these ("ope, BMW, no way he'll let me go").
My shadow theory is: people overwhelmingly associate roundabouts as being better for pedestrians because of an unrelated knock-on effect: Cities which invest in roundabouts generally have more money than cities who don't, and cities which have more money are also likely to independently invest in dedicated pedestrian infrastructure, like dividers and rail trails. Roundabouts might be safer for pedestrians, because I'd guess that many pedestrian deaths are the result of cars running red lights, and that's just physically more difficult to do with a roundabout; that's only one component of "better" though, and even then I struggle to find a better data-backed source for that intuition than "some academic or city planner said they're safer for pedestrians".
Where I grew up, if a car is stopped in front of a pedestrian crossing, not stopping next to it gets you a pretty hefty fine. Every few years the police have a big public service announcement campaign about it, then focus on giving fines for it in some of the more risky locations.
Roundabouts are better when they are not the multi-lane monsters US tends to build.
https://www.carmel.in.gov/government/departments-services/en...
https://www.iihs.org/topics/roundabouts (ctrl f pedestrian)
A "roundabout" is a kind of intersection. One characteristic intersections generally have is that traffic should not stop upon entering the intersection. The picture you've linked is a circular road, with a half-dozen or so individual signaled intersection along it.
While roundabouts do have center islands, sometimes even with sidewalks, pedestrians generally are not expected to ever walk there. For the same reason as outlined in the previous paragraph: getting pedestrians safely to the center island would require interrupting the flow of traffic inside the intersection. Walking by some roundabouts with built-up center islands, I sometimes wonder how long someone could camp out in the thicket without ever coming within 20 feet of another human on foot; the islands are no-man lands, very unlike the satellite image you link.
I can reasonably envision a pedestrian signaled roundabout, where the crosswalks at the entrances and exits to the roundabout are all signaled. I have not, to my memory, ever seen one. And, again, I think the reality of their rarity speaks to my original point that pedestrians are, as in all US traffic design, always an afterthought.
There are actual roundabouts in DC, but typically they don't have names.
I was in London area in the early 90's and they had stoplights at roundabouts that were high volume. If I remember right, a person told us the lights are only used at certain times of the day.
And when I'm sitting at a red light with a bunch of other people, complaining to myself about how inefficient it seems because there is currently no cross traffic, I have to remind myself we are helping people downstream from the intersection to enter the road.
The on-ramp meters provide the service of prioritizing drivers from distant suburbs or out of town at the expense of those who dwell in nearer suburbs. To put a different spin on it, they slightly disincentivize short trips on highways.
> There would be times when you could not get on to the highway at all, when traffic was heavy but not so heavy that it slowed.
The flowing traffic makes space for the merging traffic, which also takes up any existing gaps. The merging traffic ideally zipper merges at traffic speed.
Traffic speed is correlated to gaps between cars. If merging gets tight, gap size is small, traffic is therefore going slower. If merging is impossible, then gap size is tiny, traffic speed is approaching zero, and both the onramp and highway are a traffic jam.
It is the accumulation of new traffic that slows traffic speed. I think you are implying traffic speed remains a fixed constant and merging traffic just waits. Instead there is an interplay.
Which comes to onramp lights. That is just a rate limiter to avoid the highway from going into an extreme bottleneck. The highway capacity is the bottleneck, so adding more traffic to it freely makes it exponentially worse. It is not at all about incentives but preservation of traffic flow.
The idea of prioritizing distant suburbs falls a bit flat when considering the evening counter commute. The idea does hold somewhat for morning rush hour as on ramp meters do prioritize the speed of existing traffic. Regardless, the rate limiting is the purpose.
Though, why are we talking about on ramp meters when the article is not about on ramp meters?
> This doesn't happen because everyone understands that those merging onto the highway are moving fast and running out of space, and prudence dictates that you slow down to make space
Merging traffic should be matching, not exceeding traffic speed. Existing traffic should _speed up_ to facilitate merging. The merging cars merge behind, not in front when cars are side by side. If a merging car does not get space, then they either slow, stop, or shoulder drive. A person might slow to facilitate a merging car that is well ahead of them. Generally that right hand lane should start opening gaps between every car to facilitate a zipper merge. If side by side, the traffic that has space, the existing traffic, should speed up to make space behind and to maintain traffic speeds. If they can't speed up, then there is overall congestion, gap size shrinks and traffic speed slows..
The coolest implementation I've seen was in Los Alamos. The sensor was way ahead of the intersection so by the time I got to it, without slowing down, the light changed. This was more than 25 years ago.
This format was really fun, it's a shame Google discontinued all its code competitions.
You can have speed limit signs, but nothing controls speed better than physical infrastructure -- in city cores, things like pedestrian islands, bulb-outs, raised crosswalks; throughout, roundabouts in place of four-way stop and light-controlled intersections.
There's an absurd amount of stop signs in the US and Canada -- practically every intersection has at least one. They're mostly superfluous, and could be done away with, replaced with proper education for drivers, and yield signs where priority is ambiguous. Stop signs can serve a purpose where the view is obscured, and the driver genuinely needs to come to a complete stop to evaluate if they can proceed safely. When they're planted everywhere, they mean nothing.
Add on top of this a fine system that's proportional to the driver's net worth and the kinetic energy of the vehicle at time of infraction, and we could be getting somewhere.
I honestly believe driving should be strongly discouraged, made more painful and more expensive, to push people away from it as a primary mode of transport. Decades ago car companies sold us a nightmare of a "freedom dream" and destroyed so much in the name of profits, it's unforgivable.
Help me out here. You want fewer or no stop signs and instead to rely on “proper education for drivers”? What is the problem you’re aiming to solve with that?
The way stop signs are used in North America basically assumes that everyone is incredibly stupid and cannot be trusted with a dull stick. Perhaps that has something to do with this individualistic bend and a hatred for collective well-being and the general fabric of society; I don't know.
Cherry on top are speed bumps placed 30' before a stop sign. That's just plain moronic. A speed bump (or a raised crosswalk) would work perfectly well with a stop sign -- both signage and physical infrastructure indicating the need to stop. But, of course we won't build it that way!
The idea comes from intersections in southeast Asia for example. It's a massive free for all with higher safety rates than signalled intersections. In the face of uncertainty, drivers slow and focus on situational awareness and negotiation. That this is actually safer, is counterintuitive.
Which goes go the larger thesis, over reliance on signaling. An idea that if you blindly follow the rules, that you'll be safe. It's the reason you should hesitate look both ways when proceeding from an intersection - to make sure someone else is not barreling thru, and that you're not blindly proceeding on the green to then get t-boned.
"Safety" rates in southeast Asian nations are not the same thing as collision rates.
In most of the southeast Asian nations I've visited, people don't care all that much if their cars get dinged up, scratched, or sideswiped at intersections. It's just part of the gradual demolition derby of life.
That won't fly in high-income places where people value how their cars look.
unrelated but i love the poetry in this sentence :)
But "Google uses AI to..." is such a frustrating way to frame it!
Sure, there are likely some traditional ML models and techniques involved in this work. It first launched in 2021 so it's unlikely there were any generative models (which, let's face it, is what many people in 2024 assume when they see "AI" mentioned).
It's not inaccurate to call it AI, but it adds about as much information to the story as saying "Google uses algorithms/computer science/data analysis to..."
The more interesting component here isn't the "AI", it's the vast amount of road usage data. Google has been able to collect from prople running Google Maps for navigation. Though maybe their PE team don't want to emphasize that as much!
It's a product marketing article, not a technical article or research paper. The details of the type of models and optimizations they use is probably a trade secret, but you can likely make an educated guess. As you say, it's mostly about applying known methods to existing data and identifying a product and market for the result (or the other way around).
4 years ago, they would have used the term "ML" (which nobody outside of tech grokked anyway). The new term is AI, and it also is the term that the public broadly understands.
In a lot of people's views, outside of our tech bubble, when they think "AI" they are thinking things like. General Purpose AI, ChatGPT, or this almost magic thing that this single thing is what is doing all of this.
When the reality is thats not the case for the vast majority of this. It isn't like they are feeding this data into Gemini or another LLM and just tell it, process this data. Or its some magic single "Google AI" that is suddenly doing all of these tasks. But that is the public perception that is being painted for non technical people. Which is dangerous, it just further fuels this idea that ChatGPT and similar is a general purpose AI.
Its a standard ML model that we have been building for many years. Whats even more concerning is ML and Machine Learning are no where in this article.
Yeah its marketing, but its frustrating seeing how much we are now just calling everything "AI" as this magic term.
Somehow everything is AI now... a washing machine with a water hardness sensor that adds some extra washing detergent if the water is hard? AI! PIR sensor activated lights? AI! Everyting is AI!
People understand “math” as well or better than they do “AI”, but the perception is way different
“Doing math” makes you think that google is filled with curious little scientists doing some good will hunting stuff on a whiteboard to help you out.
“Using AI” makes you think that google is using its proprietary algorithm to do something mysterious that’ll probably benefit google and/or have unintended consequences for your town.
> Sure, there are likely some traditional ML models and techniques involved in this work. It first launched in 2021 so it's unlikely there were any generative models (which, let's face it, is what many people in 2024 assume when they see "AI" mentioned).
I don’t think you have a correct mental model of the history of the field. “Generative” is not a very clear technical term and most ML from this era is not going to be “traditional”/classical - it’s almost certainly still DL
There is a classic distinction between generative and discriminative models but it does not fully map on to what people mean when they say “generative” nowadays
I don't think that applies to this Google project: no human could perform the work needed to optimize traffic lights to reduce stop/starting, the dataset they would have to consider is way too large.
s/frustrating/profitable
2ish years ago, this would just be another ML tool that Google released. Would be interesting, but it wouldn't be "we did this magic AI thing" as if ML didn't exist a couple years ago.
It is honestly getting really frustrating seeing this drastic shift in branding from ML to AI, that outside of LLM's (mostly) is the same thing we have been doing for several years anyways.
I happen to know a few googlers socially from my days in SFLUG and BayLUG, so I can sometimes get a note to the appropriate product manager or engineer through the friends-of-a-friend network. But going through the front door has never worked.
I believe that Google Maps is perfectly fine in most of the bay area. It's generally acceptable for roads that a google street view car has driven down. But pretty much a lost cause for other roads.
Corner me at a conference sometime and I'll tell you about how google maps sent us four-wheeling through eastern california fire roads (a dirt road that collapsed after we turned around at a critical junction) or told me to get on and off the highway the first time I drove from Tacoma to Seattle (what should have been a simple 45 min drive with traffic turned out to be a 2 hour slog fest because I didn't have enough local knowledge to realize maps was full of itself.)
While it's great they no-doubt are providing an internship for two or three PhD candidates, I think they may want to fix their data before thinking AI will improve the experience.
The AI does not know the data you're training it with is garbage. If yoy do it right, you may be able to spot anomolies in the data or auto-cluster bits of data, but if you train any sort of CNN on garbage data, you're going to get garbage out.
So while this may be great for people commuting from the GooglePlex in Mountain View to the Google facilities in San Francisco, and it //may// help people traveling along I-5 in California, I fear the garbage aspect of their geo data will not be magically solved by adding AI.
I remember learning 10+ years ago in college that (I think) Paris was using AI to dynamically optimize traffic lights to manage congestion. It was called machine learning back then, of course.
The oldest paper that Google lets me find today is from 2017 (search has a recency bias): https://ieeexplore.ieee.org/document/8122189
edit: found one from 2014 – https://www.dot.ny.gov/divisions/engineering/technical-servi...
Edit: Found better sources!
Here's a paper from 1990 about real-time modelling and monitoring of traffic patterns in Paris https://www.researchgate.net/publication/317769123_Modelling...
And a paper from 2015 that explicitly mentions using cameras for traffic detection: https://hal.science/hal-01491597/document
Many intersections don't have traffic sensors.
The systems I heard about in college used cameras. Lots of CCTV in cities like Paris :)
Wish I remembered what it was called so I could find sauce
Edit: Found better sources!
Here's a paper from 1990 about real-time modelling and monitoring of traffic patterns in Paris https://www.researchgate.net/publication/317769123_Modelling...
And a paper from 2015 that explicitly mentions using cameras for traffic detection: https://hal.science/hal-01491597/document
When I heard about this in college the systems used the existing CCTV surveillance network to analyze traffic.
I am familiar with Peter Sanders' 2009 talk on Fast Route Planning (https://www.youtube.com/watch?v=-0ErpE8tQbw).
Seems like it might be useful for testing hypotheses about how proposed road projects might impact traffic flow.
> We build an AI-based model of each intersection, including its structure, traffic patterns (such as patterns of starting and stopping), light scheduling, and how traffic and the light schedules interact, and then we also build a model of the interaction between traffic lights. Based on this model, we develop AI-based optimizations and then provide recommendations to city engineers via the Green Light interface. As an example, we might identify an opportunity to coordinate between intersections that are not yet synced and provide a recommendation around the timing of the traffic lights so that traffic flows more effectively along a stretch of road.
It's not that it's inaccurate to call it "AI", it just doesn't actually tell us anything useful about what they did.
I don’t see what the problem is with that
https://spectrum.ieee.org/amp/your-navigation-app-is-making-...
Then people have to solve their problem of simply not being able to commute in their own city by finding alternative routes. Which they would do without Google Maps anyway. Maps is just making it easier. So I disagree fundamentally with the generalization that these apps are "making it worse" just because they are making it convenient. The alternative is that people just don't do what they want to do because they don't know how, which isn't "better."
I would rather blame the people not solving the problem (governments and voters) than the people who are (companies and apps) in this instance.
The article you shared also has a false premise. If the main roads are congested, and people diverge from that road and congest other roads, there is still more overall traffic flow. And this isn't Europe, where we have thoughtful city planners in every major city. The city planners aren't fixing the bridge near my house and 4 lights within a short distance, or the static 50-second timed lights on the alternative route any time soon. I'm not breaking their genius plan because their plan wasn't genius in the first place.
In fact, project greenlight addresses some of the problems this article you linked complains about (the fact google and city leaders were working separately)
This seems to be about optimizing flows globally, and there's a feedback loop between Google maps and intersection timing. For instance, if the majority of traffic flows east/west through an intersection, then you could expect more time is given to that flow than north/south. This would minimize global drive time, but it wouldn't be locally optimum for people that drive north/south through that intersection.
Google Maps – this tiny residential street looks empty, let's send a million cars through it.
That said, anything less than a motorway/A-road (trunk road) seems to get treated the same, whether it's a wide well marked road or a track with grass growing in the middle and foliage scraping the sides of the car.
Do they use the same road weighting sets in all countries?
I think the algorithm should probably weight more highly the time taken to reverse back to a passing point, and trying to turn right out of a tiny road into a busy main road behind a queue of ten other cars doing the same (which it fooled me into doing a while back).
I agree with you that it doesn't seem to really know the difference between a less-major A road and a farm track. This makes navigating over Dartmoor interesting.
And the roads are the wrong colours. </grumble>
The number of times that I've had to miss exits or turns because google re-routed without sufficient lead time has caused me to stop using it for driving navigation.
It's no wonder you see crazy people swerving across 3 lanes to make an exit these days. I blame google maps navigation for a lot of unsafe driving I see on the road today.
0% of that is Google's fault.
Everyone knows that maps will recalculate the route if you miss your exit. "Crazy people swerving across 3 lanes to make an exit" is purely caused by selfish, dangerous drives who don't care about endangering themselves or the people around them.
I've mostly given up trusting Google directions and just use it for overall route planning but then doing navigation myself while driving.
All the young folks in our family who use their phone for navigation (all of them) are helpless if they lose connectivity or have to figure out a route themselves.
We've lost something.
Nope, they are traffic!
The word pedestrian, or bicycle does not exist in this article whatsoever. Traffic lights are a flawed, but useful tool for traffic calming. If we optimize for reducing the amount of red lights cars will run into, this will fundamentally increase speeds on roads, and increased speeds, equals increase pedestrian, cycle, and car deaths.
This is all conjecture, but it seems like the key indicator that Google is providing two cities is reduction in stop time. If that is their key metric, while also not looking at other things like bicycle stop time, or pedestrian wait time; we will be optimizing for average car speed indirectly. That is a bad thing inside of cities.
The researchers would have to be very naive to omit this from their model, so I am going to assume that they did take it into account. But I agree - it's disturbing not to see it mentioned at all.
Let's go with an absurd example: If you multiplied by 5 both the length of red and green stages, I would expect much less accidents per day, but it would obviously also be much more frustrating to move around (and possibly more accidents if the amount of infractions increase because of the frustration). If you divided them by 2, I would expect much more accidents of all kinds.
>Today, Green Light is live in over 70 intersections
I'm not even sure if this is a good production result to know what has been accomplished.
It's almost like its weighing a low average speed with green driving and a long string of stop lights will lower the average speed.
This requires a lot of specialized infrastructure to be run. I really like the idea of cities investing in measures that make small positive improvements for many people. However, for Google this seems like a side project, so I have my doubts that it will be around in 5 years. Not exactly the kind of thing I want to back with long term infrastructure dollars.
By using an offline GPS app, you no longer have the need to coordinate with a centralized server. In exchange, you will lose traffic data. There may be some app that gives traffic with a level of privacy loss that meets your needs, but generally the data collection compromises your privacy, and the data is valuable... Anyway, OSMAnd is a popular F-Droid app I remember hearing mentioned.
But, if you don't need traffic, then you could theoretically navigate with all your radios off, so you wouldn't be talking to towers. The experience will probably be quite slow, even with all the navigation constellations available now. The increasing incorporation of L2 and L5 GPS may improve that, but generally you're asking for niche performance, so I wouldn't hold my breath.
I think people typically leave data on and prefer to just buy prepaid phone plans anonymously, paired with some combination of VPN usage and not using the phone for anything with a login, hoping to break the chain of tracking between the phone and the user's identity. It's probably enough to stay out of geofence warrants, at least.