Is model 3 ready for future self driving?

[OPRCE]

I am wondering: should I (can I) pay for FSD now so that I get the HW3 upgrade? I love EAP as it is with my model 3. Would my experience possibly be significantly better with HW3? Can anyone share thoughts on what HW3 is and how it might change the user experience, for example with regard to highway driving or other aspects?

1. AFAIK you can still buy the off-menu FSD option post-purchase, but for how long that will remain the case is questionable. Also, it is most likely to become more expensive as time progresses.

2. It is up in the air at the moment as to what HW3 will bring, or rather when.

3. Musk has predicted that it *should*, with the addition and refinement of future software, suffice to bring a HW2.5 vehicle up to L5 FSD autonomy (no driver needed) by end of 2020 with regulations catching up by 2023, whereas Tesla has given no official timeline.

4. They will inevitably release the new FSD features in stages, possibly starting from summer 2019, still leaving the driver legally responsible for supervision of the vehicle, same as for current L2 AP, until all features can knit together into a unified FSD and it has been road-tested for several billion miles in the fleet. At some intermediate stage along the way to L5 FSD, EAP may be declared e.g. a Highway-qualified L3/L4 system where the driver can be disengaged and the owner's insurance picks up the tab for any errors in the AI's judgement [Musk has declared that Tesla aims to avoid all liability forever].

5. If this vision all works out flawlessly, by 2023 you will be able to send your car to autonomously pick up passengers and ferry them wherever. The car will in theory do everything a human driver would need to do on public roads, in a statistically safer fashion.

6. However, Tesla and flawless are words rarely found in the same sentence, as to date their execution of plans has been, to put it charitably, uneven.

7. It remains to be seen whether the raw processing power of HW3 combined with their software design will be up to the task of L5 FSD. If it transpires not, a >=HW4 will become necessary.

8. Then the fitness of Tesla's HW2.5 sensor suite must be called into question -- problems here include:
A) The forward radar is of such low resolution that at highway speeds it cannot distinguish a fire-truck parked in your lane from the guardrails along the roadside, so the signal is ignored and the system relies solely on visual recognition, which thus far has resulted in several smashes and deaths.
B) The forward cameras may be partially blinded if facing into a low sun. Without sensor redundancy, e.g. LiDAR or radar which function reliably across the whole range of operating speeds, this spells trouble.
C) The cameras in general have a relatively low resolution. Competitors are planning to use ones x4 better, hence can detect detail at greater distances.
D) This is all the more relevant due to the fact that Tesla lacks rear corner radars for reliably detecting the speed of motorway traffic closing from behind. It seems to me quite imaginable that e.g. on the German Autobahn, where cars overtake @280kmh [174mph] in the rain, a Tesla on FSD (or EAP with unassisted lane-change) may not easily recognise such a vehicle silhouette coming shrouded in spray without lights in daytime and pull out into its path, thus causing an horrific smash.
E) The reversing camera cannot be counted on for redundancy in this case either, as it is set at the wrong angle, is short range and highly susceptible to distortion from water droplets.

9. So there can be many's a slip 'twixt the cup and the lip. IMHO at bare minimum the radar sensor will definitely need to be upgraded. Cameras are 50/50.

I am not sure what you mean here? What do you mean by "EAP only cars" or "just for EAP"? I thought HW3 was basically a new chip to enable additional enhancement of autopilot?

10. In theory it is to enable the jump from current EAP to FSD but you may be half right -- new cars from April 2019 coming pre-equipped with HW3 but without the FSD option will most likely show significantly improved performance in EAP over those running on HW2.5, despite the fact that Musk has proclaimed HW3 is not necessary* for EAP. Technically that will be correct, in that EAP will still [be made to] run on HW2.5 at the limit of its physical capabilities, just comparatively badly I should think.

(* i.e. this excuse was conjured up in an attempt to evade having to retrofit masses of vehicles at company expense if the owners do not buy the FSD option, although it is possible they will be sued and have to do it anyhow)

11. Whether you should now buy FSD depends on how much you enjoy a slow-paced technological adventure versus affording the risk that it never pans out as promised.

 

[MaryAnning3]

Thanks a lot OPRCE! Really informative post.! If I may ask about your earlier post, in which you alluded to problems when a lead vehicle moves aside to avoid something (occluded object problem?), I wonder how serious that is presently. There seem to be only a handful of stories about collisions arising from that circumstance. I definitely think about that when I am following with EAP on, but I wonder how significant you think that problem is in terms of overall driving safety issues with EAP?

 

[OPRCE]

Thanks a lot OPRCE! Really informative post.! If I may ask about your earlier post, in which you alluded to problems when a lead vehicle moves aside to avoid something (occluded object problem?), I wonder how serious that is presently. There seem to be only a handful of stories about collisions arising from that circumstance. I definitely think about that when I am following with EAP on, but I wonder how significant you think that problem is in terms of overall driving safety issues with EAP?

You're very welcome!

I consider the occluded stationary object problem extremely serious and unresolved as of latest software v9.2018.48.12.2. In my own case (link in signature) I would have been completely wrecked or killed if not paying strict attention in that rare moment which happens about once in a year of motorway driving, i.e. leading vehicle cuts out to reveal stationary vehicle in lane up ahead.

There is also evidence it happens at much lower speeds, to wit ~50mph in the wild
and controlled tests from June 2018 confirm v8.1 still suffers this phenomenon.

In official EURO-NCAP tests of mid-October 2018 [late v8.1 sw AFAICT] Tesla HW2.5 still failed the cut-out test at 80kmh [@1:05 in video], though the FCW did at least sound off:

The biggest problem is allowing oneself to be lulled into a false sense of security by all the happy-clappy PR from Tesla, thus increasing the likelihood of being distracted in precisely the moment when most needed to save oneself.

This combination is ATM the most dangerous risk from AP, because it is the most treacherous. Tesla have yet to admit the inadequacy of the radar sensor and have done little or nothing to fix the issue since Jan 2016:

Until an independent test centre (such as Thatcham Research, UK) demonstrates on video that this weakness is definitively overcome in all lighting conditions, then IMHO it is imperative to operate under the assumption it remains a deadly trap for the unwary.

Certainly I will not bet my life on any statement from Tesla or Musk and recommend scrupulous attention and follow distance = 5 or upwards at 80mph on motorways to allow more space for any mishaps.

 

[SeventyFix]

Original Poster: Yes, I agree with your opening statement. When shopping for the Model 3 in December 2018, I was clearly told that the car was 100% ready for full self driving when that software and testing was completed. Furthermore, I was told that full self driving was currently being tested on the fleet of Tesla trucks and that the download (and functionality) would be sent to me automatically and for free (I purchased every option that was available for the car (on Tesla.com in late December 2018), with the exception of the white interior).

I was also told other things that turned out to be inaccurate.

 

[Knightshade]

Original Poster: Yes, I agree with your opening statement. When shopping for the Model 3 in December 2018, I was clearly told that the car was 100% ready for full self driving when that software and testing was completed. Furthermore, I was told that full self driving was currently being tested on the fleet of Tesla trucks and that the download (and functionality) would be sent to me automatically and for free (I purchased every option that was available for the car (on Tesla.com in late December 2018), with the exception of the white interior).

I was also told other things that turned out to be inaccurate.

If you bought every available option then what they told you was accurate except for the fact you'll need to undergo about a 30 minute plug and play chip swap that will be free (and possibly done at your home/work by a mobile ranger, though that's not yet clear).

 

[OPRCE]

I wonder how significant you think that problem is in terms of overall driving safety issues with EAP?

Quite apart from the question of sensors, using EAP without forethought has a few other "features" likely to contribute to collisions:

1. longer stopping time due to car maintaining speed/acceleration while driver moves foot to brake during an intervention, compared to regen deceleration beginning immediately upon lifting foot from amps pedal if driving manually. Let's say starting speed is constant v=80kph [50mph], time moving foot to reach brake [a part of overall reaction time of 1.5s on average] is the same in both cases at 0.25s, and regen scrubs 5kmh off speed in that time. In EAP, from the time you hit the brake, the car still has to slow from 80kmh whereas driving manually it would slow from 75kmh. As kinetic energy of a moving vehicle is a function of its mass times velocity squared, this exponential relationship can make a surprising difference to the braking distance, in this example ~5m on a dry road.
How to Calculate Braking Distances

2. this delayed reaction factor is worsened if the driver has acquired a "learned waiting", in expectation that EAP should react, before making any emergency intervention. Another quarter second is easily imaginable.

3. EAP always tries to drive up to the speed limit, even at night on winding local roads, which is in general too fast to stop for the unexpected.

Here is an example of a M3 in AP doing 60mph on local road at night while driver is slightly visually impaired with a migraine and an injured deer is sitting in the road [WARNING, PETA-Disapproved!]:

Given Tesla's logistics hell that damage may well mean 6 months of stasis until spare parts get delivered and it could as easily have been a seated person driven over undetected [there was no FCW, although set to early, or AEB] which would entail some legal questions.

 

[AlanSubie4Life]

Poor deer. Poor Tesla. Glad no one was hurt with that amount of ramping over the deer.

Seems likely that a vehicle not on EAP would have been going more slowly at impact - if nothing else, the driver probably would have removed some pressure from the accelerator in response to seeing the deer, even if they did not hit the brakes.

Definitely agree with the idea that EAP could make this sort of collision worse. When I use it on the freeway I always worry about road debris - because I would be paying much closer attention, by necessity, if I did not have EAP on, I would be more easily able to avoid such things. Even though I am paying attention, the attention is not the same, as much as I would like it to be. And then there is lack of smoothness resulting from taking over autosteer, in the event of taking evasive action. It’s one of the things I like least about having to take over steering if I do not have time to disengage Autosteer - it fights enough I end up jerking the car around at freeway speeds in a manner disconcerting to passengers.

Getting a computer to drive is a REALLY difficult problem, even if you have an extremely elaborate and complete sensor suite. Seems virtually impossible with the existing set of sensors but hope I am wrong. At least the FSD folks will get whatever they need, assuming the self-driving problem is solvable. Computers have a huge advantage of effectively instant reaction time, but can’t react to what they can’t see.

Looks like that car might still be driveable. Hard to know if the radiator is damaged - probably is. Maybe it can be patched up and driven around until the parts are ready, to reduce rental expenses and inconvenience. Always feel bad for a Tesla owner with a vehicle that is not drivable but not a total loss.

 

[SeventyFix]

If you bought every available option then what they told you was accurate except for the fact you'll need to undergo about a 30 minute plug and play chip swap that will be free (and possibly done at your home/work by a mobile ranger, though that's not yet clear).

Glad to hear it and I hope that you are correct. The self-driving aspect was a large factor in my purchase. If successful (and I believe that it should be), I see this as a major transformative technology of my lifetime (up there with the internet, MRI, gene editing, etc). Imagine a future with no (or wildly reduced) drunk driving deaths, far lower traffic fatalities, etc. I would imagine that many of us in the United States pass at least one of those small roadside memorials put up by families of people who have died at that very spot in a traffic accident. Imagine a future where one could sleep, work or read a book in the back seat while the car drove itself. This was once science fiction - now it is absolutely plausible (yes, agreed, we're not close to that - driver needing to be prepared to take over at a moment's notice, etc). But it could be a reality one day.

 

[sperkin]

Quite apart from the question of sensors, using EAP without forethought has a few other "features" likely to contribute to collisions:

1. longer stopping time due to car maintaining speed/acceleration while driver moves foot to brake during an intervention, compared to regen deceleration beginning immediately upon lifting foot from amps pedal if driving manually. Let's say starting speed is constant v=80kph [50mph], time moving foot to reach brake [a part of overall reaction time of 1.5s on average] is the same in both cases at 0.25s, and regen scrubs 5kmh off speed in that time. In EAP, from the time you hit the brake, the car still has to slow from 80kmh whereas driving manually it would slow from 75kmh. As kinetic energy of a moving vehicle is a function of its mass times velocity squared, this exponential relationship can make a surprising difference to the braking distance, in this example ~5m on a dry road.
How to Calculate Braking Distances

2. this delayed reaction factor is worsened if the driver has acquired a "learned waiting", in expectation that EAP should react, before making any emergency intervention. Another quarter second is easily imaginable.

3. EAP always tries to drive up to the speed limit, even at night on winding local roads, which is in general too fast to stop for the unexpected.

Here is an example of a M3 in AP doing 60mph on local road at night while driver is slightly visually impaired with a migraine and an injured deer is sitting in the road [WARNING, PETA-Disapproved!]:

View attachment 370050

Given Tesla's logistics hell that damage may well mean 6 months of stasis until spare parts get delivered and it could as easily have been a seated person driven over undetected [there was no FCW, although set to early, or AEB] which would entail some legal questions.

Thanks for sharing. Btw, the camera on the Tesla makes it seem the deer is really close but in reality the driver had more time to see it and react. This driver wasn't even paying any attention. Also, the camera is darker than it actually is. Driver could have seen it well before in camera view.

What I wish Tesla do is have sensors seeing objects under 4 feet tall. Car doesn't care about ladders, buckets, chairs, young children on the ground.

 

[Derek Kessler]

What I wish Tesla do is have sensors seeing objects under 4 feet tall. Car doesn't care about ladders, buckets, chairs, young children on the ground.

That's well within the realm of machine vision capabilities. It's just not something Autopilot can do today.

 

[sperkin]

That's well within the realm of machine vision capabilities. It's just not something Autopilot can do today.

I didn't think about that. Hope it can detect debris where the car can't clear like anything over 6 inches.

 

[jasdelta]

1. AFAIK you can still buy the off-menu FSD option post-purchase, but for how long that will remain the case is questionable. Also, it is most likely to become more expensive as time progresses.

2. It is up in the air at the moment as to what HW3 will bring, or rather when.

3. Musk has predicted that it *should*, with the addition and refinement of future software, suffice to bring a HW2.5 vehicle up to L5 FSD autonomy (no driver needed) by end of 2020 with regulations catching up by 2023, whereas Tesla has given no official timeline.

4. They will inevitably release the new FSD features in stages, possibly starting from summer 2019, still leaving the driver legally responsible for supervision of the vehicle, same as for current L2 AP, until all features can knit together into a unified FSD and it has been road-tested for several billion miles in the fleet. At some intermediate stage along the way to L5 FSD, EAP may be declared e.g. a Highway-qualified L3/L4 system where the driver can be disengaged and the owner's insurance picks up the tab for any errors in the AI's judgement [Musk has declared that Tesla aims to avoid all liability forever].

5. If this vision all works out flawlessly, by 2023 you will be able to send your car to autonomously pick up passengers and ferry them wherever. The car will in theory do everything a human driver would need to do on public roads, in a statistically safer fashion.

6. However, Tesla and flawless are words rarely found in the same sentence, as to date their execution of plans has been, to put it charitably, uneven.

7. It remains to be seen whether the raw processing power of HW3 combined with their software design will be up to the task of L5 FSD. If it transpires not, a >=HW4 will become necessary.

8. Then the fitness of Tesla's HW2.5 sensor suite must be called into question -- problems here include:
A) The forward radar is of such low resolution that at highway speeds it cannot distinguish a fire-truck parked in your lane from the guardrails along the roadside, so the signal is ignored and the system relies solely on visual recognition, which thus far has resulted in several smashes and deaths.
B) The forward cameras may be partially blinded if facing into a low sun. Without sensor redundancy, e.g. LiDAR or radar which function reliably across the whole range of operating speeds, this spells trouble.
C) The cameras in general have a relatively low resolution. Competitors are planning to use ones x4 better, hence can detect detail at greater distances.
D) This is all the more relevant due to the fact that Tesla lacks rear corner radars for reliably detecting the speed of motorway traffic closing from behind. It seems to me quite imaginable that e.g. on the German Autobahn, where cars overtake @280kmh [174mph] in the rain, a Tesla on FSD (or EAP with unassisted lane-change) may not easily recognise such a vehicle silhouette coming shrouded in spray without lights in daytime and pull out into its path, thus causing an horrific smash.
E) The reversing camera cannot be counted on for redundancy in this case either, as it is set at the wrong angle, is short range and highly susceptible to distortion from water droplets.

9. So there can be many's a slip 'twixt the cup and the lip. IMHO at bare minimum the radar sensor will definitely need to be upgraded. Cameras are 50/50.



10. In theory it is to enable the jump from current EAP to FSD but you may be half right -- new cars from April 2019 coming pre-equipped with HW3 but without the FSD option will most likely show significantly improved performance in EAP over those running on HW2.5, despite the fact that Musk has proclaimed HW3 is not necessary* for EAP. Technically that will be correct, in that EAP will still [be made to] run on HW2.5 at the limit of its physical capabilities, just comparatively badly I should think.

(* i.e. this excuse was conjured up in an attempt to evade having to retrofit masses of vehicles at company expense if the owners do not buy the FSD option, although it is possible they will be sued and have to do it anyhow)

11. Whether you should now buy FSD depends on how much you enjoy a slow-paced technological adventure versus affording the risk that it never pans out as promised.

Where can I find off menu FSD on the Tesla website?

 

[gaM3changer]

Where can I find off menu FSD on the Tesla website?

In your account, go to manage and scroll down. The shop section will show Autopilot Upgrades.

 

[jasdelta]

In your account, go to manage and scroll down. The shop section will show Autopilot Upgrades.

Many thanks.

 

[OPRCE]

Thanks for sharing. Btw, the camera on the Tesla makes it seem the deer is really close but in reality the driver had more time to see it and react. This driver wasn't even paying any attention. Also, the camera is darker than it actually is. Driver could have seen it well before in camera view.

What I wish Tesla do is have sensors seeing objects under 4 feet tall. Car doesn't care about ladders, buckets, chairs, young children on the ground.

I'm thinking that, in the absence of LiDAR or RaDAR reliably picking them up, a forward-facing infra-red camera would be very helpful for safe autonomous night driving, so as to detect wildlife, drunkards in dark camo and unlit cyclists on the road.

 

[sperkin]

I've been very critical about Tesla's Autopilot but it's getting better and better with every update. There were situations where I get surprised that AP will see the cars now. Usually AP will track with the car merging on the freeway to a point where they will collide if the other driver doesn't speed up or slow down. Now AP will actually slow down and let the other car merges on the freeway.

However, AP is still very robotic and stiff. It doesn't speed or slow down very smooth. Changes lanes like a 15 year old behind the wheel. Hope all of this will get fixed in 2019. Very excited for FSD for onramp off ramp. Also, hope they remove the nag to touch wheel.

 

[Atari2600]

Until they add some sort of auto cleaning to the side and rear cameras the system will not be ready for self driving.

EAP currently handles additional lanes poorly. When driving north on i71 right before West Lancaster/Jeffersonville an additional lane opens up on the left side so the road transitions from two to three lanes. My car never stays in the current lane and for a brief time splits the two lanes then quickly changes into the leftmost lane without any signal. I've grabbed the wheel dozens of times driving through that area hoping it would trigger something on the back end but it still does it on 2018.50.6. How hard is it to just follow the right side of the lane where the lines do not change?

 

[NuclearPowered]

Until they add some sort of auto cleaning to the side and rear cameras the system will not be ready for self driving.

EAP currently handles additional lanes poorly. When driving north on i71 right before West Lancaster/Jeffersonville an additional lane opens up on the left side so the road transitions from two to three lanes. My car never stays in the current lane and for a brief time splits the two lanes then quickly changes into the leftmost lane without any signal. I've grabbed the wheel dozens of times driving through that area hoping it would trigger something on the back end but it still does it on 2018.50.6. How hard is it to just follow the right side of the lane where the lines do not change?

I've noticed a marked improvement of this issue in my daily commute in 50.6. 1 lane splits right in the middle to two on the interstate. The car used to pinball wildly and randomly pick one. It now always chooses the left without any hesitation or swerving. NOA has gotten worse to the point where I have stopped using it.