Cheap 'n' simple sign trickery will bamboozle self-driving cars, fresh research claims Eggheads have taken a look at previously developed techniques that can be used to trick self-driving cars into doing the wrong thing – and found cheap stickers stuck on stop and speed limit signs, at least, are pretty effective. They also learned of a specific phenomenon where the systems "memorized" signs, so that when they …
"spatial memorization design" could also be problematic. The fact that a road was a 40mph limit last week doesn't mean you can assume it's still a 40 limit just cos you've not seen a sign. "It's always been a 40, your honour, how was I to know the limit had changed" isn't a defence you can rely on on court.
Something like that may have been done close to where I live. A section of road that had a 50 limit suddenly has no 50 signs at all.
The problem with this is that in theory when passing through this section from one direction the speed limit is 30, yet if you are travelling in the opposite direction the speed limit is 40!
As far as I'm aware in the UK they do not have different limits according to which way you're travelling. So either someone has screwed up an official speed limit change, or someone has covertly (and illegally) removed the 50 limit signs on this section of road.
I even tried searching the websites for the local councils, but was unable to find anywhere to find out about speed limit changes, or where to report this anomaly.
It will probably not surprise some to learn that I live in Wales!
If it's a council road then there should be a Traffic Regulation Order describing any change to speed limits etc.
For 'trunk' roads I presume it would be Highways England or Transport Scotland etc.
I have an odd one on my way to work.
Dual carriageway separated by trees, i.e. 2 lanes - trees - 2 lanes, about 1 mile in length.
On the side on my way to work the speed limit is 70, it's a pretty straight road on that side.
Most people tend to drive it at 55 or less in both lanes for some reason which is a PITA.
Coming home from work it starts out at 70 for a few hundred metres then drops to 40 AFTER a sharp bend, from that bit it's a tiny bit bendy so stays at 40.
Obviously everybody does 50+ since it only changed a couple of years ago and previous speed limit is ingrained.
Would have made far more sense to change the speed limit before the bend since all the accidents were on the bend itself.
Only example of this that I've ever seen but obviously there must be more if you've got similar.
I have an issue for example the A41 at Hunton Bridge. The road is a dual carriageway and well lit by a system of lighting where the lamps have been placed no more than 200 yards apart. The speed limit in the area is National Speed Limit Applies. Which according to the definitions would make the speed limit 30, except everyone drives it as if it's a 70.
There are countless examples of this up and down the country.
Another particular annoyance is the "Pass Either Side" sign, which is actually defined as "Pass Either Side to Reach Same Destination". It gets used all over the place willy nilly to mean "Don't crash into this bit where the road divides."
As for Germany: For the triple it is often 100 on the right lane and 120 for the middle and left lane. Happens more often is the speed signs are electric, but some with static. For dual carriageways I've never seen it for fixed speed limit signs, but for electric ones. Oh, and when there is construction the right lane might be 60, and the left 80, but rare.
Round here - near Potsdam DE - we have not only roads which vary anywhere between 50-100 depending on how likely someone thinks a tree might leap out at you, but also limits which are dependent on both time _and_ day. And what you're driving - trucks may have lower limits in the village than cars do, for example.
So the automation not only has to read the limit sign but also the small print underneath it, and get both right, and then hope its clock is right too.
I got the impression from the article that the "sign disappearing time" was fairly short, as in: if a sign was seen and visibly of it was subsequently lost, then it was assumed that the sign was still there. This would make sense when, for example, queueing at a junction with an observed stop sign which then gets obscured by a van coming alongside.
I agree after, say, 10 mins this assumption gets iffy given that the sign may be temporary.
This is what made the TSR more susceptible to "appearing attacks": the stickers only had to fool the TSR briefly and it thought it had gotten a glimpse of a sign.
A few years ago I was driving a hired Passat, with a fancy-pants "smart" cruise control which was supposedly linked to a front-facing camera that could read speed limit signs, but clearly had a "back-up" database, woefully out of date.
One section of road had been changed from derestricted to 40 about 10 years before. The Passat got to that point and deployed the loud pedal, despite the 40 repeater sign and the stream of cars in front of me.
My TomTom satnav still has loads of TSR from many years ago, because it's one of the models from which they with drew map updates rather than fixing a GPS rollover bug. It claims the time is always 12:00 but apart from that - and the out of date maps - it still works fine. Even when the maps were supposedly up to date it tried to take me across a field next to Schiphol Airport and then across a bridge in Utrecht which had been closed for four years.
> changed from derestricted to 40 about 10 years before.
A private school on the main road here went broke 10 years ago. I know cuz I went to the tag sale. I don't know when the signs and lights were taken out, but it's been years. Never were any kids in sight of the road (very small enrollment and set well back). Still my nearly new Toyota has to panic-light "SCHOOL ZONE" when I pass that point. The internal maps have updated several times (usually when I am lost, so leaving me lost longer) so it ought to have the recent data. I fear, being a private ex-school and a public roads department, nobody can talk to each other.
I live in a cul-de-sac. You can get into my road from both ends as a pedestrian, but any vehicles can one use one end of the road.
Until last year, whatever GPS/Mapping app I used would not direct anyone via the end of the road that is pedestrian only even if asking for pedestrian directions.
There is main road running by the end of the road. Most Satnavs would not direct anyone to go past my road at this end, tending to take a longer route to avoid a bit of road that according to the various mapping companies does not exist.
Also, down the road, the residents of a few streets have put up official looking signs advising Truck drivers to ignore their GPSs and turn back.
This is what concerns me about self driving systems. They can be excellent at navigating, but if they have out of date maps, they may well have problems. At the moment, you, as a driver, can take over, but how long will that last?
Round these parts there's several locations along derestricted single carriageway roads* where a long term temporary speed restriction has suddenly appeared on a 50m stretch.**
*60mph, but only if you don't mind the risk or wiping yourself out against an incoming bloody great tractor), where there
**Long term access to underground cable installation works for Vattenfall wind farm long distance DC feeder
My Rav4 RSA system seems to remember old speed limits for roads. When joining a road with a 60km/h limit that used to have an 80km/h limit, the car thinks it's still an 80km/h limit. It sees the 60 sign before the junction and briefly sets the new limit but when joining the new road, it jumps to the old 80 limit.
After reporting this to Toyota with an example of where this behaviour happens, there seems to have been an OTA update for that specific junction but it doesn't solve the general issue.
Annoying behaviour for an RSA system and I doubt it's any kind of mitigation against a speeding fine but these systems (like it or not) are legislated in Europe to improve road safety. Manufacturers should have a duty to implement them in a manner that doesn't result in this kind of scenario which arguably has entirely the opposite effect.
A long way to go before I would trust a self driving car...
You can sort of see where this might head. Things like mandatory signs might end up containing beacons that passing cars can read passively, in addition to the visual cues for human users.
And I'm betting that given Tesla's increased reliance on visual sensing for its stuff, as opposed to some other manufacturers, it makes them more vulnerable here.
Actually a passive RFID like device embedded in the sign which encodes the speed/times/vehicle restrictions that approaching vehicles could scan might not be the daftest idea.
Although in the current shit storm of insanity even a chocolate teapot would rank among the less daft.
I once got trapped at a crossroads where the four "no entry" signs had been rotated to suggest there was no way out. I presumed straight-ahead would be the legal route and proceeded slowly in case I was wrong.
I have always wondered how a self-drive vehicle would handle that situation. Hopefully the same way.
Where I live some traffic lights are always at red during dead of night, unless a sensor detects a vehicle waiting for the light to change and then, eventually, goes green (this happens for intersections between main traffic axes and secondary ones).
If the waiting car isn't close enough to the sensor, or if the sensor malfunctions (something not that uncommon), then red never goes away.
Humans know what to do in a situation like that, self-driving cars would be stuck there forever
Regina (and many other cities) in Saskatchewan have sensors on the traffic lights to detect emergency flashing lights and change the lights to green if an emergency vehicle is sensed approaching. So at night as you approach, assuming no other traffic, you can flash your headlights and if you get the timing right, the lights change to green. You're not supposed to do it, and the local plod take a dim view of it, but useful.
"Where I live some traffic lights are always at red during dead of night, unless a sensor detects a vehicle waiting for the light to change and then, eventually, goes green"
I used to drive early morning southbound on the A38 toward Bristol, through a set of lights that would be set happily on green whilst nobody was about, then quickly change to red when it sensed me approaching. They got ignored because visibility was good.
Another set of lights on a pedestrian crossing in Weymouth would reliably turn red if you approached them above the speed limit, but not if you didn't.
Will just sit there. I Was just at a conference in SF and took Waymo.
It was in the right turn only kerb lane and needed to move over to go straight but it won't push into traffic unless there is a large enough gap so was forced to go around the block
A limo was parked blocking the hotel drop off so it sat at the junction for 3 cycles of traffic lights with people honking at me until I gave up and got out
Due to spatial memorization, if you trick a car with one of the "appearing" stop signs, will it always (wrongly) stop at that spot in the future? Is that information automatically spread to other vehicles?
One false stop sign on a 55-mph road would wreak havoc.
You would hope it would have the sense to know that there aren't typically stop signs in the middle of motorways.
However, here in the land of continual road maintenance, it will have to interpret the vague gestures of bored teenagers working as flaggers around construction sites.
No, but on uk motorways, they can close lanes with a red X above them. It's possible all lanes can show this. And besides, the car doesn't have "sense", it has code. What would a human do if it came to a "stop" sign while on a motorway?
Stop?
Proceed with caution?
Ignore and continue?
I'd suggest each answer will get about 33% of the vote. Good luck programming that!
"Stop? Proceed with caution? Ignore and continue?"
And the chap who drops a U turn.
How a driver might respond to all Xs above the lanes would also depend on where the lanes were headed. Only the insane or suicidally stupid would proceed into a tunnel. (Or even a bridge or floodway.)
I've encountered TSRs on motorways before, where the next gantry is blank. Human drivers seem to interpret a blank gantry as the speed restriction is lifted rather than National Speed Limit Applies or End of Speed Restriction. An unlit gantry could be faulty or in the process of commissioning. they're nearly impossible to see at night on unlit sections anyway.
Ever been past a terrible accident where the 4-lane has been chopped to a 1-lane? They'll station two guys with stop/go signs on poles at either side to continue at least some traffic passing.
This reminds me of the scene in the museum attic in Thief of Time (Pratchett) with the sign that said keep right over an arrow pointing left (or vice versa).
I suggest putting a 1 in front or a 0 after the printed numbers on speed limit signs.
And then you end up with a crash as occurred in Lake City, FL in the 1950's. An intersection had a wooden stop sign erected in the late 30's and after it rotted away in the late 40's the locals continued to treat it as a stop sign. An out-of-towner blasted through the no-longer-actually-there stop sign in the mid 50's and t-boned a local. Originally written a ticket for ignoring a posted sign, the ticket was tossed because the picture of the intersection showed no sign...
There's a little bit of Spain call Llivia which is entirely surrounded by France. The Spanish road N154 connects it to Puigcerda, about a mile away in Spain proper. On the way it crosses the French RN20. These days it passes over a bridge, but untll fairly recently it was a crossroads, at which both routes claimed priority.
-A.
As I understand it, the "memorization" is just a short term thing that the car "forgets" once it has passed where it thought it saw a sign, or after an appropriate time-out. Something like this would be necessary to deal with issues such as the sign getting hidden, partially or completely, by shrubbery or other things around it as the car approached it.
In other words the sign may be hidden intermittently as the vehicle approaches it by small trees, other signs, bus shelters, etc., so the vehicle has to "remember" that there was a sign there during the periods when the sign is obscured instead of forgetting it instantly when the sign is momentarily obscured. A human driver would know do this without being told, but the AI system has to be explicitly programmed to take it into account. This is one of the real world problems that any such system has to be able to deal with in an imperfect world.
Some of the image trickery systems seem to work by using patterns that confuse the image recognition system as it passes from one image cell to another due to parallax as the vehicle moves. In other words, the stationary sign appears to "move" from the imaging system's perspective due to the motion of the car, and the patterns applied by the stickers confuse the image recognition system as they slide across the field of view.
I suspect that in the long run if self driving cars become something other than a novelty found on a very small number of cars authorities will install passive RF markers in appropriate places to supplement signs for cars to use instead of relying only on visible signs. If they don't then the self driving feature would become unreliable for half the year due to snow obscuring signs and road markings. A human driver can deal with a stop sign that is covered with snow simply by noting that there is a sign in roughly the spot where a stop sign could be expected to be given the surrounding terrain context. An AI may not reliably make that decision however.
The graphics card is not needed for image generation or printing, but for picture editing.
And even then, an 8086 CPU, DOS, HP DeskJet 500C (if you can find one) is enough. You can use a 286-CPU, Windows 3.1 286-mode (or GEOS) and a whooooping 1 MB of RAM if you want to go fancy.
They are taking Arthur generating anti-patterns for the vehicle's sign recognition system. These are visible in the photo of a grid of modified so signs. Some of them tell the system it's not a stop sign, others obscure the signs' message when placement.
I might be stupid, but aren't all these self-driving features supposed to have cloud backup of their data? The "spatial memory" could be as simple as a maps database with signs logged. So when a new one shows up, it could self-update the shared map provisionally until/unless it is confirmed by other observations -- other personal cars having minimal weight of confirmation, a dedicated robo-mapping car (e.g. Google Street View) having highest confirmation weight.
If every car was connected and shared a database, it might become the most accurate map ever... or become a major single point of failure that dooms hundreds to injury or death. Either way, no different than relying on our handbrains/pocket computers to navigate for us already!
(Of course, none of this matters for their testing on top of a parking garage, where maps normally don't -- shouldn't? -- cover.)
that would require a mandatory database for all construction sites, accidents and whatever temporary stuff which require redirection.
The result would be: "No sir, according to the data you are in the middle of the road. It is your mistake to put your bedroom is there. The data is flawless."
In the USA, sure - try that in Europe with some of the tiny roads we send HGVs down and it wouldn't last five minutes.
Besides, I can't imagine American truckers having anything to do with autonomous trucks - they seem to thrive on active discomfort as part of their jobs. See this ongoing series from Bruce Wilson where he takes a European-spec Scania around the USA and all the people who look at it or drive it say how wonderful and advanced and futuristic it is compared to their American boneshakers.
I mean it's obvious stuff like US trucks having leaf springs on the front instead of full air ride, or not having rear axles which lift under light loading to save on tyre wear and fuel economy.
Some years ago I read a post on a website that suggested that bands of highway guerrillas might roam the country, altering or spoofing road signs (as is happening with GPS now) and any other objects that 'smart' cars use for navigation, causing chaos on the roads, before disappearing until their next attack.
I found myself milling around South Island (New Zealand) recently in a hire care that was equipped with a bunch of those new fangled bells and whistles. The three main systems were speed limits, automatic headlight dip and lane departure and the car had a small head up display to inform the driver what it was doing. I used the system a lot, primarily because I couldn't figure out how to turn it off (there's no single "lobotomize" switch) and found it was easy to get used to. I was surprised to find out just how well everything worked. Speed limit detection was perfect, only occasionally being confused by the signage at the end of road works. Automatic headlight dip worked perfectly both for oncoming traffic and when you caught up with someone. Lane departure was a bit iffy, it never quite got the hang of NZ's single lane bridges. Overall, everything worked really well, it just seemed to me overkill, a novelty with just yet more sensors and actuators to fail expensively at some time in the future.
Everyone seems to be quick to find flaws with automatic driving systems, criticizing every small trick and detail as 'the end of the world as we know it' while overlooking the effect that similar sign trickery would have on an 80-something senior with barely passable eyesight. Most drivers wildly overstate their driving ability anyway so roads and their signage has to be designed for not the sharpest tools in the box, be they computer of wetware. The only guarantee going forward is that the computer is going to get better than me (if its not already).
"I found myself milling around South Island (New Zealand)"
Possibly one if the easiest places for FSD to work. The resident population is pretty sparse and the traffic density even on the major roads low. I reckon peak hour in Dunedin would have to be gazetted to be noticed. The sealed roads while narrow are generally in good condition and well signed.
The greatest hazard pre COVID was the number of Chinese (PRC) tourists hurtling around mountain roads† at suicidal speeds having clearly obtained their licences from a cereal packet.
A beautiful and serene part of the world worth visiting but I lived there until 13 so am probably biased.
† eg Te Anau to Queenstown - one petrified clown skidded to a stop a mere metre or so from a 500m precipice.
> Maybe meat drivers are just much better.
Many years ago I relocated within the United States and applied for a driver's license from my new jurisdiction. My awareness of traffic regulations was (lightly) tested with a series of photos of driving situations, each with a multiple-guess "What should you do?" question. One picture comes to mind whenever I contemplate "self-driving" cars: The photo was of a four-way intersection with a traffic signal suspended over the middle and showing red--but there was a uniformed policeman in the middle of the street waving me forward. Would a Tesla or Waymo or Cruze obey the signal, or would it follow the policeman's directive? Would it follow the instruction of a non-uniformed school crossing guard? How about a random Joe Jerkface in some kind of emergency?
"Scanning officer's badge number ... Verifying badge with police database ... HTTP 503 ..."
My local authority once planned a complicated set of coordinated roadworks, with multiple sometimes-overlapping road closures and diversions. Oh-oh, you're way ahead of me. Knowing about some bad things that had happened recently, as a local Councillor I publicly asked the Highways chief to ensure that there would be no diversions down closed roads this time. He laughed, "Of course not! I can assure you that won't be happening again." When I encountered the inevitable, I posted a pic on facebook and notified the Highways guy accordingly. My, that soon had it fixed!
But what if I had been an autonomous mechanical dildohead? [Stop sniggering at the back!]
Human-piloted moon landers have so far had a 100% success rate while self-guided probes have failed to stay in one piece and right side up in half the cases. [1] And that's on the moon where there's no other moving traffic, pranksters with stickers and tape, or a variety of objects in all shapes and colours to confound the unwary algorithm.
And yet we expect self-driving cars (produced within consumer budgets rather than using aerospace-grade technology) to do much better? Really?
[1] Granted, this is based on a very small sample size. Still, the trend is there...
Making self-driving work properly is expensive and difficult (maybe too difficult). There will be lobbying to introduce laws to make doing stuff to confuse robocars a crime (see jaywalking). Then "doing stuff to confuse robocars" will be widened to include things like walking or cycling. Buying politicians has long been more cost-effective than solving problems
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
