Anyone else concerned about fog without radar?

[ajdelange]

No concerns at all. A human may drive faster than the fog should allow but FSD is smarter than those people and will drive at a speed it can safely react to.

In fact when a fog is encountered FSD announces that it cannot see the lane demarcation lines and shuts off.

Sponsored

 

[jerhenderson]

That helped me understand what they are up to. In essence he said that the radar is very good at measuring range, velocity and acceleration but that the cameras are just as good. He also said the the algorithms that turn the radar data into tracks are not so good and loose lock fairly frequently and that this is something they could fix but if the radar doesn't really allow them to do anything better than they think they can do with the camera why waste resources on fixing the tracker.

He also mention wide vertical beamwidth problems which mean that the radar cant separate car from an overpass. I'd say they picked a pretty poor radar for this appliction! And yes, I have had my car scream bloody murder when there isn't another car in sight. He again said that this is fixible (it only takes more vertical aperture) but again not worth doing as they don't think they need a second sensor. Seems they have thrown everything into neural nets and let fusion, autonomous integrity monitoring .... slide.

I shall, nonetheless, miss radar. I have been excited by it since I first set foot in one (SCR 584 - I had a complete set of the TM and the dipole feed from one as a kid) and worked as a radar engineer for a few years at the outset of my career. That's why I laugh at some of the assertions made here. Please don't take them seriously. I'm sure I'd be wowed by the capabilities of a modern radar. I'll bet they haven't lost the ability to resolve in range 0 doppler though. In fact I'll bet the ambiguity functions are much closer to the ideal "thumback" that we strove for those 50 yrs ago.

your vision is also very good at measuring range, velocity and acceleration. FSD is better because it's free of your distractions and humsn assumptions.

 

[jerhenderson]

In fact when a fog is encountered FSD announces that it cannot see the lane demarcation lines and shuts off.

that's good too then..... a good warning to the driver that circumstances warrant extreme caution. one day it'll learn I'm sure.

 

[ajdelange]

Regarding the large number of comments in support of radar the sad reality is that when you have multiple sensors of differing types you are going to have to decide how to listen to them.

Yes. That's an important part of system design. The AI guy recognizes that the sensor integration is flawed. He then goes on to say that the single sensor is so good that it's npt worth bringing the system design up to snuff.

In fact, it is probably better, in the long run, and when using an ML-based system, to use just one sensor type.

That is almost never true. What is true is that the additional sensor, or sensor type, comes with a cost (it must be integrated properly and a cross checking integrity monitoring system must be part of that) and a benefit. If the OEM does not feel the benefit is worth the cost, then he abandons the second sensor. That's the case here.

 

[ajdelange]

your vision is also very good at measuring range, velocity and acceleration.

No, actually it isn't. Try an experiment. I'll be particularly interested to hear how good you are at estimating acceleration.

FSD is better because it's free of your distractions and humsn assumptions.

FSD is better because it has 8 cameras as opposed to my two and can compute faster than I can. But it is fascinating that even 2 cameras is enough to feed the Kalman filter that is clearly running in our brains in the background. Now my wife doesn't have one or if she does it isn't tuned properly and it is fascinating (and terrifying) to observe the way she drives.

 

[tmeyer3]

What system allowed the Tesla to hit the broadside 18 wheeler? The news said the car detection system was not able to discern the gray trailer in the hazy sky so I assumed a visual camera system but other news said the sensors were looking at the road and looked under the trailer once it stretched across the road indicating more of a radar looking down at the road. Either way the company may be looking to find a system more accurate. This Tesla video from a couple years ago is clearly radar based and predicted the crash without seeing it. I want this in limited visibilty, not so I can drive faster but so that my CT can see beyond what I can see.

I know I'm going to get roasted for saying this, but even without the radar in use, my Tesla can still see 3+ cars ahead of me. No, I'm unsure how, so don't ask me. What I do know is that there are a LOT more cameras on my car than what you see in videos or what is displayed to you. The front alone (mid-dash) has at least 3, but I've only ever seen 1 displayed or record. These capture different frame rates, measure distances, etc. I've also been told that Teslas are built to see 360 degrees around the car and process all of it continuously, which isn't a stretch with modern computing.

I realize that on the surface of the argument "radar vs visible light", it is intuitive to think "oh, the visible light cameras could never see that coming as well as the radar." But they really, really can. At extreme ranges, yeah radar would win, but not on the ground, too many emissive objects around.

In a Tesla, you can see what they call the "mind of the car". It's a little virtual world that is displayed to you that illustrates what the car can see around you. You don't have to believe me, time will tell all! But take a watch at this:

This guy is not driving, and no radar is involved.

 

[LDRHAWKE]

I’m going to put this on top of my Cybertruck so I will be safe in a fog.

 

[GnarlyDudeLive]

Honestly, I think Tesla should keep using radar not for FSD or autopilot but for anytime there is manual override of either one of these systems. The data collected at this moment transferred to Tesla from the radar camera could provide some value to the computational backend AI in making better decisions for the visual only cameras next pushed updates. That is if they are not already doing this.

If our eyes are considered good enough to drive with I see no reason that a multi-camera visual only system can't meet or exceed human vision with a good AI. Zero distractions and far more camera angles and the ability to digitally calculate speeds of multiple simultaneous objects and trajectory's precisely are all things capable with plain ole visual only cameras. This all assumes the vehicle is getting zero information from other nearby vehicles which could enhance the whole process 10x more easily.

As far as attempting to do the full speed limit while under adverse weather conditions even if your own vehicle could do it safely, I would be very fearful you could spook a lot of other human drivers when they detect your AI driven ride approaching quickly and causing an ancillary accident via knee jerk action or panicked response.

 

[Ogre]

None of this matters.

Tesla has made it clear they are not using radar. Seems like they are doing pretty good with this whole self driving thing so far. While they aren’t as far along as they thought they would be…nobody is ahead of them either.

I pay Tesla to figure this stuff out and let them worry on the details

 

[tidmutt]

You should watch Dr. Karpathy's presentation on Tesla's FSD, explaining why fusing together vision and radar was less successful than vision alone.



This video is why I'm a believer in the vision-only paradigm, although I agree that Tesla probably pulled radar prematurely due to supply-chain issues.

This guy works in traffic safety engineering… he doesn’t think Tesla pulled the Radar due to supply chain or cost:

https://www.quora.com/Why-is-Tesla-...er/Frank-Zucco?ch=99&share=48b567ae&srid=ny7J

 

[ÆCIII]

On one hand I am a little puzzled Tesla seems to be completely throwing away the radar implementation. While I understand possible issues of false indications causing phantom braking or other behavior, I would think they could prioritize it as a secondary backup input to visual information and possibly react to it's information only if visual input is blinded (i.e. fog, extreme weather). But apparently that's exactly what they've been doing and have found the incidence of reliable usage minimal. Yet I think if it was perfected more, it could add lot of benefits during bad weather and poor (optical) visibility. It would be interesting to see visualized representations of what the radar is giving to the 'mind of the car' compared to the cars discerning of visual information from the cameras.

But on the other hand it puts all the focus on AI processing of visual information learned without depending on any secondary radar input or 'crutch' - so it should result in more AI/visual refinement. I guess we will 'see' where that takes us.

But IMO if depending solely on visual input from cameras, Tesla needs to do something more about the water/fog or road splatter build up on cameras by implementing some kind of air nozzle or mini wiper to clean them whenever the view is obstructed. Currently when driving in rain or snow, we're sometimes getting errors saying cameras are blocked or obstructed (mainly pillar or side repeater cameras in my experience), and the rear backup camera often has to be wiped of water when driving in rain. I've actually pulled over at rest areas in bad weather to wipe off cameras obstructions so that Autopilot would stop giving errors.

If Tesla is going 'all vision', then I think they should be seriously addressing obstructions to their cameras to assure their maximum reliability. If Tesla needs real world visual input to 'learn' and evolve its AI, then visibility even in poor weather conditions seems like an important concern. If FSD works reliably (only on 'clear sunny days'), then how will regulators be able to allow it?

 

[jhogan2424]

your comparison isn't valid as FSD will keep a safe distance from that car and react faster than you ever could. Radar isn't needed. The majority of accidents are caused by human error - FSD doesn't have that weakness.

Your comparison isn’t valid as I never mentioned FSD. I did not order FSD because I do not want it at this time. I would rather have radar than FSD for the extra 10k.

 

[rr6013]

I currently drive quite successfully in fog without radar. I expect this to continue.

Straight ahead foggy progress notwithstanding, fate in ambient fog or frosty fog is dependent on the “other guy” seeing your car.

Depth perception is the first sensory input to be corrupted in fog. Looking out into the nether fog the human brain works overtime mapping the 2D scene to the brain’s 3D knowledge base. SO if the other guy can see your car, he may not be able to interpret your place in space thus misjudging distance. Worst accidents in fog are simply people pulling across your lane of travel, knowing you are there, not able to recognize that the image they see is closer than it appears. T-bone!

Snow blindness is the second sensory input to be corrupted in fog. Looking out into the nether fog the human brain works overtime mapping the 2D scene to the brain’s 3D knowledge base losing its ability to resolve color space. White colors and grey colors become indistinguishable, edges fade and longer the thicker the fog, the human eye loses all integrity.

THUS while radar, computer enhanced AI and TeslaVISION cameras do exceed human capability. That doesn’t make you, personally, one iota safer in fog. NOR does all the excess technology in the world make your car safer on the road in fog, frosty fog and smokey conditions. There remains the “other guy” sitting in his car stopped at his driveway or side-street looking to enter the roadway or cross. He’s human.

 

[TheLastStarfighter]

It's human nature, but the fuss over radar is only because many do not understand the problem of FSD. The challenge is not in the ability of the vehicle to sense its surroundings, but in the ability of it to interpret them. Once the programming can accurately interpret what it perceives and pilot using this data, we will have FSD. The roadblock is not lack of sensors.

As such, radar is pointless. It's neither precise nor accurate. It can tell you there is an object, but not the objects dimensions. It can tell you speed, but only within 10% accuracy at best (always fight a speeding ticket within 10% of the posted limit!) Cameras provide infinitely more data, and 8 cameras are more than enough to provide dramatically more information than a human with two cameras (eyes) can ever take in.

As a human driving, you would never depend on radar. You wouldn't (or at least shouldn't) go faster in fog or snow because you have radar. In fact, radar doesn't work in fog or snow. But regardless, you're not going to make driving decisions without 100% visual verification of your surroundings. The FSD system should do the same. If the cameras don't see something, it shouldn't depend on the less accurate radar. If it can't drive using the cameras, it shouldn't be driving at all.

 

[rr6013]

What system allowed the Tesla to hit the broadside 18 wheeler? The news said the car detection system was not able to discern the gray trailer in the hazy sky so I assumed a visual camera system but other news said the sensors were looking at the road and looked under the trailer once it stretched across the road indicating more of a radar looking down at the road. Either way the company may be looking to find a system more accurate. This Tesla video from a couple years ago is clearly radar based and predicted the crash without seeing it. I want this in limited visibilty, not so I can drive faster but so that my CT can see beyond what I can see.

Interesting.

This inside lane situation my defensive driving got my car out of harms way. Same situation though I’d picked-up danger through the car ahead w/s. Braking was not safe. I knew my right lane was open and clear from previous scan. I switched lanes avoiding colliding with car ahead. Car behind me had more distance to brake.

IDK radar latency. I would bet my money right now that TeslaVISION reacts quicker in this situation shown and mine described. Would it change lanes? Doubt its Situational Awareness is there, yet.

Impressive video well worth the safety coming soon to a Tesla near you.

Sponsored