FSD may require a hardware upgrade...

[3Victoria]

Tesla seems to have lost at least 1 year by going to Tesla vision and dropping mobile eye.

Probably not. Two different systems, and AP1 has a distinct limit to its functions. Each system has a learning curve. AP1 is at the levelling out phase of its curve, AP2 is in the early stages. Mobileye's sensors and pre-processors have limitations, and they would have had to go with their newer systems which themselves are undergoing their own development curve, so even if they had stuck with M'eye, they would not be further along. Keep the faith.

 

[lunitiks]

You can't say Lidar is better than radar. Or better than cameras. You just can't. Not without specifying what it's better than the other sensors AT. Look, Lidar can't read lane markings. Lidar can't read traffic lights, and it sure can't read speed limit signs. Furthermore, it can't see through fog, smoke, dust, rain and snow as radar can.

In essence: Lidar sucks. Except it doesn't. In fact it marvels at HD mapping. It's a master at generating the most high precision, super perfect maps of the cars surroundings.

Except it isn't. Enter a bit of snowy mowie, and it's f***ed.

 

[AmpedRealtor]

I can see Tesla supporting the central board upgrade but when you're talking about a different sensor suite that is completely different. Does anyone think the sensor suite won't change in the next 2 years?

I'd like to think that the version 2 sensor suite will have much more life than the original suite, however it's missing some hardware that almost all other autonomous vehicles in testing are using - multiple radars and lidar. Tesla's solution has only one forward-facing radar and that's it. I would think that to achieve true FSD the car will need multiple radar units to see beyond the ultrasonic sensors.

For instance, a rear-facing radar would be necessary in order to provide rear collision avoidance, or to at least prepare cabin occupants for an impending impact by tensioning the seat belts and issuing a visible and audible warning. Front and rear corner radars may be required in order to provide complete 360-degree coverage and a certain level of redundancy for the cameras, more confidence in identification of pedestrians and other "soft" objects, etc.

I'm just speculating, but I don't think Tesla has magically figured out a way to eliminate sensors that other manufacturers have deemed necessary for their autonomous technologies (i.e., lidar and additional radar units).

 

[electracity]

You can't say Lidar is better than radar. Or better than cameras. You just can't. Not without specifying what it's better than the other sensors AT. Look, Lidar can't read lane markings. Lidar can't read traffic lights, and it sure can't read speed limit signs. Furthermore, it can't see through fog, smoke, dust, rain and snow as radar can.

In essence: Lidar sucks. Except it doesn't. In fact it marvels at HD mapping. It's a master at generating the most high precision, super perfect maps of the cars surroundings.

Except it isn't. Enter a bit of snowy mowie, and it's f***ed.

So you are saying that LIDAR produces a superior 3D map only 99.5% of the time compared to the technology Tesla uses? I doubt any of the current LIDAR cars on the road today are doing emergency braking for overhead signs.

I don't know if Tesla needs to use LIDAR, I don't think they know either. What they do know is that they could not afford to put LIDAR in the model 3 in 2017/18 as standard equipment.

Initially the Waymo/Google type FSD car will be corporate owned with expensive hardware. It can have many types of sensors.

 

[Ohji]

I just want to know what happens if there is a catastrophic hardware failure while the vehicle is in autonomous mode. Does anyone know of any instances where AP1 hardware suddenly failed during autopilot use?

I think that the AP2 hardware may have full autonomous capability, but the redundancy factor is what concerns me. If the system senses it is failing and can give even a 1-2 second warning for the driver to take over, that is fine (assuming there is a driver, lol). Modern chips are pretty reliable, but I could see the cooling system failing, for example, causing the PX2 to shut down to prevent overheating. I would imagine that there would be enough warning to alert the driver to take over, but what happens if AP is navigating a sharp curve on an undivided road and the chip fails in the middle of the turn? Does the car keep its trajectory or does it run off the road (or worse, into oncoming traffic). If it fails suddenly, the second or so it would take for the driver to realize that there was a problem and then take control back might be too long, yes? Of course, the PX2 is a multi-chip solution, so maybe the whole board would have to fail for there to be a catastrophic failure.

This redundancy concern is a moot point if FSDC ends up being a sytem where the driver is never able to keep his/her hands off the steering wheel, but if the intent is for the driver to eventually be able to focus on other tasks besides driving (or to have the car drive itself), then the safety backup is something to think about. One other possibility -- maybe Tesla just has to demonstrate that the chance of a catastrophic failure of the system is significantly lower than that of a human having a sudden stroke/heart attack. Redundancy may not be required in that case -- the reliability just has to exceed that of a human suffering a similar "catastrophic failure."

I am going to try to limit my expectations with regards to FSDC. I will hope that it can do what was demonstrated in the video with a somewhat alert driver in the driver's seat, but I am not going to expect that I will be able to have my car drop me off at the airport and then drive itself back home (although I would really, REALLY like for it to be able to do this, lol!).

 

[lunitiks]

So you are saying that LIDAR produces a superior 3D map only 99.5% of the time compared to the technology Tesla uses?

Say what? Read my post: Lidar is superior - except NOT in rain, snow and dust. Why? Because these are conditions where the short wavelenght laser light hits millions of rain drops, snow flakes or dust particles and generates MAPS out of them. Not so useful ones.

Forget maps. Think of the amount of false positives - "ghost objects"; in essence total blindedness - you would get from a fwd facing lidar as your only or primary sensor in traffic.

To reiterate: No sensor is better than all the others. They're great at each of their specific, designated tasks

 

[mblakele]

I think that the AP2 hardware may have full autonomous capability, but the redundancy factor is what concerns me. If the system senses it is failing and can give even a 1-2 second warning for the driver to take over, that is fine (assuming there is a driver, lol). Modern chips are pretty reliable, but I could see the cooling system failing, for example, causing the PX2 to shut down to prevent overheating. I would imagine that there would be enough warning to alert the driver to take over, but what happens if AP is navigating a sharp curve on an undivided road and the chip fails in the middle of the turn? Does the car keep its trajectory or does it run off the road (or worse, into oncoming traffic). If it fails suddenly, the second or so it would take for the driver to realize that there was a problem and then take control back might be too long, yes? Of course, the PX2 is a multi-chip solution, so maybe the whole board would have to fail for there to be a catastrophic failure.

Sensor redundancy is a good idea, and it's a good idea to think about failure modes. But let's not confuse safety improvements with the illusions about perfect safety, nor about human driver safety. Autonomous vehicles will have to be safer than human drivers initially, and increasingly safer over time. But they'll never be 100% safe: nothing is.

Human drivers have some sensor and processing redundancy, and plenty of vulnerabilities. We don't like to think about these things, but what would happen in your hypothetical situation if a human was in control, and had a stroke? I'm not sure how often that would happen, relative to the computer failure you posit. Stroke and automobile accidents. - PubMed - NCBI shows that it happens often enough to be studied. Human drivers are subject to many kinds of sudden failure: heart attacks, narcolepsy, migraines, etc. Even minor problems like a sneezing fit or a stray eyelash can be risky when you're behind the wheel. Let's not even get started on distracted driving.

Opinions vary on when we should accept autonomous vehicles as "safe enough": anything from "as safe as human drivers" to "ten times safer". But nothing in life is 100% safe.

 

[Reeler]

I think regulators are going to require 360 degree redundancy on the sensors. I think Tesla should have gone with 3 more radars to accomplish that. At least one rear facing radar. Maybe they could even drop the ultrasonic parking sensors in favor of more radar. Solid state lidar would probably fill the gap nicely.

Maybe v3 hardware will have a chance. Not sure why Tesla cheaped out with less processor than Nvidia recommended and far less sensors than the competitors seem to think is necessary. Perhaps planned obsolesce with everyone with v2 hardware upgrading to the new v3 when the regulators say no.

 

[Viking1]

Autonomous vehicles will have to be safer than human drivers initially, and increasingly safer over time. But they'll never be 100% safe: nothing is.

I think this is key. It will not be perfect. Perfect should not be the enemy of good. Its a numbers game. Is the car driver safer then your average human? Or can it drive better then your superior driver? I've heard Elon talk about the accident rate drop 2X, 10X, whatever it was. I think they see it as a numbers game. Maybe not the regulators. Or the public. That may be an issue.

 

[LuvMyS]

FIRST PICTURES - Tesla Autopilot 2.0 ECU (Nvidia PX 2)

Thanks .. have you seen a tear down of the new AP2 cars?

 

[lunitiks]

Thanks .. have you seen a tear down of the new AP2 cars?

This

 

[⚡️ELECTROMAN⚡️]

I wonder if they built in some upgrade-ability this time not just for the processor but for more sensors if needed. Elon seems awfully confident that this hardware will take us to full autonomy. I can't even see out of my back up camera half the time. I would imagine the four side cameras are going to have similar issues. I don't think my car will ever be fully autonomous in the rain unless it has more than just forward facing radar.

 

[lunitiks]

Well they did plan for "corner radars" at one point. Can't remember where I read it but it was b4 the AP2-reveal.

Lidar I dont think we'll see, ref. Elons discussions about this. At least not until its alot cheaper.

Cameras (actually all sensors) will have the issue of nature clogging them up somehow.

What is needed, are some super smart, auto actuated wiper-defogger-washer-thingy in front of the sensors

 

[Matias]

What is needed, are some super smart, auto actuated wiper-defogger-washer-thingy in front of the sensors

Simple washer fluid injector will help a lot.

 

[MorrisonHiker]

Thanks .. have you seen a tear down of the new AP2 cars?

We might do one eventually but our endoscope camera didn't work the last time we tried it. Also my car is at the Service Center. Maybe someday!

 

[Bladerskb]

You can't say Lidar is better than radar. Or better than cameras. You just can't. Not without specifying what it's better than the other sensors AT. Look, Lidar can't read lane markings. Lidar can't read traffic lights, and it sure can't read speed limit signs. Furthermore, it can't see through fog, smoke, dust, rain and snow as radar can.

In essence: Lidar sucks. Except it doesn't. In fact it marvels at HD mapping. It's a master at generating the most high precision, super perfect maps of the cars surroundings.

Except it isn't. Enter a bit of snowy mowie, and it's f***ed.

What are you even talking about? the criticism against tesla is their intent not to include lidar.
No one is advocating a lidar only system. Sure lidar can't read traffic light color, lanes or read road signs.
That is why you include a camera to perform those tasks. Sure lidar can't see through heavy rain, fog and snow. Which is why you also include radar in your system.

But a camera only system is bound to fail.

Camera are easily blinded by sunlight and have trouble at night, they also can't see through and are disabled by heavy fog, rain, and snow.
Radar returns only 2d information, that's why you see those false positive braking because it can't differentiate an overhead sign from an actual oncoming object. While Radar can see through snow, heavy rain and heavy fog. it is also disabled by snow and dirt build up.


In essence from your own words "Look, Radar can't read lane markings. Lidar can't read traffic lights, and it sure can't read speed limit signs. Furthermore, it can't see objects in 3d as lidar can. In essence: Lidar sucks. Enter a bit of freeway with overhead sign, and it's f***ed.

Here's volvo's sensor fusion for example:

7 radars
7 camera
1 lidar
12 ultrasonic


Here's Google/Waymo sensor fusion: Time #13m38s

7 Lidar
9 cameras
4 radars

 

[oktane]

Wow, Volvo's system appears to be much more sophisticated and redundant than Tesla's. They are showing true L5. Is this Volvo video also a phony like Tesla's? Looks like another artist rendering to me.

I think Tesla cheaped out on the sensors to keep the AP2 feasible for Model 3. I would have gladly paid another $10-20k to have a fully functional, redundant, and true Level 5 AP instead of false promises.

 

[calisnow]

Volvo doesn't have a system. They have a rolling test lab. When they have a product you can buy get back to us.

 

[Max*]

Volvo doesn't have a system. They have a rolling test lab. When they have a product you can buy get back to us.

That's a little harsh. Volvo has the Drive Me Project, where they put real people in autonomous cars. The issue is that the car can only take a predetermined route, so while it's autonomous, it's not really autonomous , but it has the potential.

Then there's Tesla, they give everyone autopilot. You can use it anywhere you please, but it's not autonomous at all.

:dunno: they both have potential for success, they both have potential for failure. We'll see who gets there first.

 

[Reeler]

Volvo doesn't have a system. They have a rolling test lab. When they have a product you can buy get back to us.

Tesla doesn't have a system you can buy either. Sure you pay for it, but you aren't delivered much more than the competition. You can buy a $25K Honda with about the same features that Tesla has actually working outside of beta.

Volvo just chooses to professionally debug their system rather then Tesla forcing their customers into a beta test. Tesla AP owners are just beta testers who paid for the privilege to participate in this grand experiment.