Tesla drivers report a surge in ‘phantom braking’
washingtonpost.com
washingtonpost.com
The article quotes: “With only one sensor type, it’s harder to be sure because you do not have the cross-check from a different type of sensor,” he said.
This is presented as though it's a clear win though. We know for example from the Uber fatality that you can also get false negatives because of failures of sensor fusion, which is a very hard problem. Requiring multiple sensors means that all of your sensor systems need to notice a danger, and you need to be able to integrate that data properly, before you can react.
So I can definitely see that there are major advantages to vision-only. But it's clear that the problem has not yet been solved by Tesla nonwithstanding Elon's inflated claims.
An increasingly complex set of if / else statements.
Is it a left turn? Is it a stationary emergency vehicle? Is it a small child? Don't worry, it's all software and we'll keep sending OTA updates until we catch every edge case.
And the loops checking like "I accelerated, it has been this much time, did I actually move 10 meters" or whatever. Then you check if something stationary shrunk by that much in size from the camera perspective...
Side tangent but some drones use this cool technology called "visual inertial odometry"
This isn't true. The principle in general is that you want to detect situations of danger, minimizing false positives and false negatives, with your pay off function probably valuing false negatives as much worse than false positives. With more sensors, the worst case scenario is that it doesn't improve the payoff function and you simply don't use the sensor information. But in the best case scenario the extra sensor can be used to massively increase your payoff - your camera gets confused by a shadow for example, but your radar doesn't detect shadows, so you can eliminate those false positives. Since the general principal to solve self-driving has been to train AI to process sensor input to make decisions, the worst case scenario is you end up with a neural net that literally just doesn't look at the other sensor data (assuming you're training your AI reasonably, if you aren't then maybe you shouldn't be working on an incredibly difficult problem with safety critical applications)
Tesla has essentially bet that they can get their optical systems so good that other sensors like Lidar do hit that situation where they offer nothing in terms of improving the pay off, but every piece of evidence we've seen so far has indicated that they haven't achieved that, and this is many years after Tesla had been confident that they would acheive it.
> all of your sensor systems need to notice a danger...before you can react
I'd be really surprised if they implemented it this way. I thought the whole pitch behind lidar was that if vision doesn't see the person walking in front of you, lidar can still see the moving person shaped thing and you can still act.
If your safety reaction is consequence-free, no problem. If it isn't (like slamming on brakes in traffic), you really want a 3rd sensor so you can do 2oo3 voting.
Suppose you have a color sensor and a form sensor and you have to identify an orange (fruit). Clearly in conjuction the two sensors will be way better, than assigning everything with color orange orange and everything round an orage.
If the signal goes from "black nothing" to "orange round" you know you went from no object to an orange, but what if your form sensor breaks and it goes from "black nothing" to "orange nothing?"
Sensor failures are independent events and need to be correctly identified no matter what.
To take Lidar and Vision, both say something about distance and form of objects. Together they can achieve better performance. If one of them fails (and failure is identified) it defaults to the other sensor only and will be somewhat worse, but should at least allow it to pull over and warn you about sensor failure.
Also failures are much easier to identify when you have a baseline reference of one or more other sensors, in isolation much harder.
This is the opposite of true. In well-designed sensor fusion algorithms, every new piece of sensor data, however noisy, helps to overconstrain the estimate. Each sensor reading helps to inform the interpretation of the other readings. If your system is designed such that more information worsens your inference, you have designed a terrible system.
What you are saying and I understand you're not saying this is the way to go but merely giving a point of view but hear me out, is that if you have two inputs instead of one, the two might disagree, which makes it harder to figure out which one is right. But if you have only one input, it can't disagree with it self so it's easier to decide.
But if the one you have is the one of the two that would be wrong, you haven't gained anything, beside being sure about it while being wrong. At least with two inputs you're able to know something is off. And usually you add more inputs so you can decide, and make a strategy depending on context (absolute majority, 50%+1, ...).
And sorry but if when having multiple input the system is confused, or needs time to correlate all the data, or whatever like that, this is still useful as again it lets you know something is wrong danger will robinson.
I think the field of engineering, with planes and spaceships and whatever, has proven how the "simplificity" advantage of single sensor is a scam that will cost lives to avoid fixing issues.
So for, not a single phantom break.
So I wonder what it is that makes it not happen for me? Probably half of those miles are in the evening, so that might be a partial explanation. Maybe the direction from which the sun is coming?
It seems the right set of circumstances convince the navigate on autopilot system that jamming on the brakes is the right thing to do.
Here are some threads about the same type of issue in VW and in Ford (the latter being on I-80): https://www.reddit.com/r/Tiguan/comments/nrkc3d/phantom_brak... https://www.macheforum.com/site/threads/phantom-braking-expe...
I've had a Rav4 for 3 years. I use the adaptive cruise control a lot on highway/freeway and I have never, ever had it exhibit phantom braking. It will sometimes brake a little longer than you expect for cars that are exiting on off ramps, or changing lanes, but it's always for a real car that's really going slower than I am.
It may well be more prevalent in Teslas though; I don't have any data on that.
Teslas have this issue way too often.
I've had it disable itself momentarily (<30 seconds) due to driving against strong sunlight (it uses visible light cameras) but aside from that the system works flawlessly.
I don't understand why more sophisticated systems that are continuously improved are simply expected to "commonly" phantom brake, but a much dumber/simpler system based on parallaxing two cameras (like the human eyeball, per the name) works reliably.
Actually I don't understanding why nobody else is using parallax in their systems? It is basically depth information for FREE using the cameras you already have. Just angle them a little, so they overlap, and then calculate the depth.
I've never had any phantom breaking when on the ACC, but I also hardly use it (I can't even remember last time I turned it on). I do, however, get "phantom" warnings from the FCW. I use quotes because, to me, the false positives are understandable. They used to happen all the time approaching my old house, where there was an S-turn with a guard rail. There were times that yes, you are driving straight towards the guard rail and the FCW would warn on that, because it doesn't account for the road turning. It's a dumb smart feature; doesn't even have lane assist. For me, it's fine. I know the road ahead. I see the curve. The system doesn't. But, because it doesn't slam on the brakes for me, it's just annoying and not a safety issue to me. If the exact same FCW was capable of breaking for me all the time, yeah, I'd have issues with it and would likely have lodged complaints with Dodge & the NHTSB.
There was no actual risk involved, and it was removed immediately with an OTA update.
Then why are they recalling them?
Should we trust their software enough to accept that it can accurately make that judgement? It's a really hard problem and they are struggling enough with the fundamentals. I don't think their software is actually at the point where they can make that assertion with any believable certainty.
No pedestrians detected, which is very different.
See e.g. the video of Tesla running over a child (mannequin, fortunately): https://twitter.com/taylorogan/status/1478802681141645322?s=...
Sadly most research in this area went out of fashion 30+ years ago.
But of course we do. It's how we throw rocks and hit what we aim at. It's how we catch things. It's how we walk around anywhere that has obstacles. We use it beyond the reach of our arms really frequently.
Basically, could you do it with one eye closed? Without stereopsis, tasks involving close-up things are harder, but farther away things are not.
If I'm running through the woods, trying to escape from a bear, stereopsis is pretty useful, because dodging trees while at a dead run is life or death.
I don't do either of those, but some of my ancestors probably did. But I play sports. I may be able to catch a ball with one eye. Getting myself to where the ball is going? I couldn't do that as well with only one eye.
So: Do I need stereopsis? No. Did the human race? Yes, it was a competitive advantage to have.
I would love to see a sports game where one team can use both eyes, and the other team only has one eye uncovered. I imagine it'd be downright funny.
This is outright false. A person with acute vision can perceive stereopsis out to 1/4 mile. Trivially, 3D movies are projected onto screens which are 10m away.
> which is why we can also understand pictures
We don't drive using pictures.
Could someone share their experience?
for the latter, long drives it helps relieve some tension in my legs to not have to be pressing the pedal to maintain speed. i cant speak for the former as i havent used them
I tend to despise most modern driving aids, but cruise control (just the very basic "maintain speed") is extremely handy and makes driving a lot easier. Maintaining a constant speed (to within 1mph) by foot over a 2 hour drive is almost impossible. It also seems to help my fuel economy.
Your foot gets physically tired pushing the accelerator for that long. Cruise control prevents that tiredness from occurring.
Adaptive cruise control is about my favorite “driver assist” technology though. Makes the constant speed changes of rush hour driving much less stressful when you use a large enough safety window. Also on open highways it is just wonderful as you basically never touch it (unlike traditional which constantly has to be disengaged for even a single other car in your lane).
Worth noting that my vehicle’s emergency braking system is distinct and active regardless of ACC status. Also I’m always ready to brake manually.
(This is purely an aside but I have a temporary soapbox so why not. I'm skeptical of more advanced forms of cruise control, even simple things like adaptive. Anything that reduces your need to concentrate tends to leave you open to daydreaming (never mind actual distractions like phones); an example of the "TSA agent problem" where staring at a screen of benign bags all day turns out to be a bad way to detect guns in hand luggage. Better I think to have nothing, absent sci-fi-years-in-the-future autonomous cars where you can sleep while they're underway. Humans are awful at passively waiting for something out of the ordinary to occur; computers are great at that. Computers suck at common sense and detecting intention; humans are amazing. The current state of self driving cars plays to both actors' worst weaknesses!)
I've driven for hours with and without cruise control. I've never felt any discomfort in my lower leg muscles because of using pedals.
Top bicycle athletes spend many hours learning how to pedal to minimize accelerations due to cyclic nature of pedaling. With a car it seems easier (all you need is to keep your foot steady, not to maintain constant torque while switching from one leg muscle to another) but I cannot do it. Cruise control beats me every time.
To expand on this: it's about maintaining a specific cadence (usually in the 100-110 RPM range). You don't change your cadence for the terrain; you maintain your cadence and change gears to match the terrain.
Constant cadence and constant speed are different goals, the first one is about biology (how to get your body to do more), the second one is pure physics. Maybe it is even "anti-biological" thing, because human's legs are designed to push, not to pull and not to create force perpendicular to a leg. I even heard something about using gears that are not exactly circular, to help an athlete to maintain constant tension of a chain.
Amateurs (me for example) do not train for constant speed, though to some extent it comes naturally with practice. But world class athletes measure energy conversion inefficiencies in tenths of a watt, so they get a special training for this.
The worst drivers are "booger cars" - they maintain the same speed as you in the adjacent lane with their front bumper roughly adjacent to your rear bumper. I call them "booger cars" because, if you try to get them to stop sticking to you by speeding up or slowing down, they just match your speed. About the only way to get rid of them is to "rub them off" by passing a slower vehicle that is in their lane, breaking their fixation on your car.
The kind of driver that especially irritates me is the one that hangs out in the passing lane, driving slowly either at or under the limit, and then when you when you try and get around them, you can't without doing +20 over the limit.
My L2 Hyundai does highway auto-cruise without issue.
In 2016, Mobileye got into a bit of a spat with Tesla, and Mobileye fired Tesla as a customer. Every other company continues to work with Mobileye's radar+optical+ultrasonic solution, but Tesla had to build its design from scratch independently.
Its no surprise to me that Mobileye, with a base of dozens-of-millions of cars, manages to do something better than Tesla (who has less than a million cars IIRC).
I sincerely hope all the progress with the FSD package, which is different from the Autopilot software causing the phantom braking, quickly leads to an Autopilot update which avoids phantom braking and is good at avoiding any obstacles.
[1] https://www.forbes.com/sites/samabuelsamid/2021/06/24/next-g...
[2] https://group.mercedes-benz.com/innovation/product-innovatio...
LIDAR should be quite for collision avoidance in conditions where it can be used. This should improve car safety quite a bit.
Cars from Lucid, NIO, Mercedes, Volvo, Polestar, Lexus. For example:
https://www.motortrend.com/reviews/mercedes-benz-drive-pilot...
https://www.motortrend.com/reviews/lexus-ls-500-teammate-adv...