Dashcam video: my Tesla smacks armored truck while on Autopilot

[Johann Koeber]

Sorry for your damage. Glad no one was hurt.

Thanks for posting. It reminds us of the need to be careful with AP.

Actually I expect it to get much better soon, but we have to be careful until proven safe in most conditions.

 

[GatorGuy]

A few thoughts and answers to your questions:

• I said it several times, I’ll say it again here. The accident was my fault. I fully admit that. I was screwing around with the screen and wasn’t paying attention to the road as I should have been. Dumb move. Lesson learned. Won’t ever happen again.

• I posted this thread for two reasons. One: for those of you who blindly trust TACC as I did, for the sake of your own driving record and the safety of others, please don’t. You can’t rely on it. Two: this is more of a philosophical debate and question: TACC is absolutely designed to slow down the car when someone moves sideways into your lane, and not just detect the car that’s in front of you at the time you turn on Autopilot. In my own experience, it sees cars moving into my lane and brakes. Just not this time. And think about it this way: it's such a basic, building block function that if Tesla can't get this right, autonomous driving not nearly close to being ready (it would be literally years and years away), and they're scamming our money asking us to pay for something that may not be available for the useful life of the car.

• So the interesting question is this: what happened here? And don’t give me that BS about what the manual says. You know as well as I do that when it says it TACC is not reliable in snow or rain, it’s talking about increased braking distance, not about the sensors being blinded by water or snow. I’m not a physics major, but I’m pretty sure radar is not affected by rain or snow. Or else you wouldn’t be able to fly airplanes in many parts of the world for most of the year.

• The audio wasn’t me talking, oblivious to my car rear ending an armored truck! It was a Bill Simmons podcast playing in the background ☺

• The AEB did beep, but I’m not sure if it would have braked because I had jammed on the brake already, just way too late.

• The truck did pull over a little down the road when it was safe to do so. Funny story: when the guard came out to inspect, I saw that he had his hand hovering over his exposed, holstered pistol. He thought I might be attempting to jack him! I apologized profusely, and he was cool after that. They turned out to be really nice people.

I appreciate you posting about your accident and with the thread. Honestly I'm a little disappointed. One of the reasons I bought a Tesla was for autopilot use in relatively low speed but high traffic situations. Very similar to this. I had been using Audi's ACC for 5 years in these exact same situations and it was bulletproof. It would never fail to identify a car entering the lane. It would be slow to speed back up until the car was fully out of your lane.

However I switched to Tesla because I thought their system would be better and to gain the autosteer capability. This isn't even an autosteer issue. It's an ACC issue where it fails to recognize or react to a obstacle in time. Every ACC system must be able to handle these situations 10000s of times.

One of the things I wonder about is the Audi system has two forward facing radar sensors with them spread on the front of the car versus just the one Tesla has. I don't know if they are long range/short range or if one covers left versus right. Maybe that gives it more reliability.

The Audi did have a longer following distance than Tesla when both are set to minimum but it didn't look like you were set to one.

DISCLAIMER: Before I get the same replies over and over, I fully understand who is responsible. But these systems are there for us to rely on. And IMO the Tesla system should be just as good as any other ACC system.

 

[mongo]

So the interesting question is this: what happened here? And don’t give me that BS about what the manual says. You know as well as I do that when it says it TACC is not reliable in snow or rain, it’s talking about increased braking distance, not about the sensors being blinded by water or snow. I’m not a physics major, but I’m pretty sure radar is not affected by rain or snow. Or else you wouldn’t be able to fly airplanes in many parts of the world for most of the year.

Some speculation:

I tried finding good data on the Bosch radar and came up with a couple potential useful bits. Some versions have dual modes, wide view and narrow view. The narrow view mode is +/- 6 degrees and goes out 150m or so. If the system was running in that mode due to speed, the radar would not see the truck at all until it was well into your lane (imagine narrow beam flashlight). With the shape of the truck, even if the radar were in wide mode. the radar return would be weak until a flat surface bounced the beam back. Similar to when you can see yourself in a shiny object, if a mirror is not pointed at you, you don't see yourself, the angle of the truck entering the lane would not provide much radar return to work with. The slow rate at which the truck moved may have caused it to be partly filtered out. A slowly increasing return may not get the response a rapid change would.

Post the Brown crash (link), Tesla made the radar the primary sensor and vision the secondary. While rain will need to be handled in FSD, it does cause issue by reducing available contrast to detect the truck visually. The light bouncing off the side is similar to the background and thus harder to detect as an object. It was also an interesting situation because there were trees in the background which could obscure the truck's silhouette initially.

 

[brkaus]

I have to share this story, seams like a good place

On a Tesla mailing list, someone said “I was so impressed with AP. It was raining so hard I couldn’t see the front of the car, but AP allowed me to drive safely at regular speed.”...

Ugh.

 

[bro1999]

I have to share this story, seams like a good place

On a Tesla mailing list, someone said “I was so impressed with AP. It was raining so hard I couldn’t see the front of the car, but AP allowed me to drive safely at regular speed.”...

Ugh.

More misconceptions of Autopilot's abilities. That guy who sent that email may be on a crash course (pun intended) with trouble.

 

[mongo]

I have to share this story, seams like a good place

On a Tesla mailing list, someone said “I was so impressed with AP. It was raining so hard I couldn’t see the front of the car, but AP allowed me to drive safely at regular speed.”...

Ugh.

Along with everyone in front of them apparently...
(or AP wasn't really needed/ doing anything)
Reminds me of the snow squall/ whiteout pile ups, as long as the car in front of you doesn't hit anything, you'll be ok, but it doesn't mean you're safe...

 

[AviP]

Its hearing stories like that that remind us
how stupid AP can be, and how much we need to stay on top of it.

https://www.navigantresearch.com/wp-assets/uploads/2018/01/LB-ADV-18-LB-pic.jpg
Navigant Research Leaderboard: Automated Driving Vehicles
The image says it all. The other companies are way ahead and while their designs look ugly, they actually seem to work.

 

[MP3Mike]

Post the Brown crash (link), Tesla made the radar the primary sensor and vision the secondary.

Minor correction, vision isn't secondary, both radar and vision are primary sensor systems.

 

[MrAustraliaTax]

In my own experience, it sees cars moving into my lane and brakes. Just not this time.

Thank you for posting your dashcam video. Others will learn from it.

You have been unlucky in learning an AutoPilot limitation the really hard way: when there is a large speed differential between a Tesla and another merging car (someone "cutting you off" - a not uncommon scenario), by the time the TACC/AP sees the merging car, it is too late.

After many years of AP use, whenever I encounter a scenario other than straightforward, well-marked lanes and other cars travelling at the same speed as me, I take over control.

AP is FANTASTIC for relieving you of the boring, mundane, routine aspects of driving.
AP is TERRIBLE at handling the unexpected.

 

[brkaus]

More misconceptions of Autopilot's abilities. That guy who sent that email may be on a crash course (pun intended) with trouble.

Same idiots drive in dust storms, snow storms, smoke without AP. Clearly they are in a hurry and don’t care about themselves or those around them.

I remember my dad, back in the 80s, once followed a semi through a major midwest snowstorm. Followed very close and all we could see was the tail lights. He was pretty proud that we didn't have to stop and get a hotel and always told the story. As he got older (wiser), he started adding "lucky we didn't all die" to the end.

 

[Electroman]

Eating oatmeal and talking on the cell phone is NOT a problem as long as you keep an eye on the road and you have one hand free to take control. Thats all.

Almost all of the issues are because the driver was not watching the road. You could be nothing but staring down the ground, and that will get you a in a bigger trouble than eating oatmeal and watching the road.

 

[dailo]

Fair enough. I do eat oatmeal instead of cereal in the car so that if I had to take over it wouldn't be as a big of a mess.

Eating oatmeal and talking on the cell phone is NOT a problem as long as you keep an eye on the road and you have one hand free to take control. Thats all.

Almost all of the issues are because the driver was not watching the road. You could be nothing but staring down the ground, and that will get you a in a bigger trouble than eating oatmeal and watching the road.

 

[sorka]

I'm having a difficult time understanding why improper use of the autopilot system has rattled your confidence in Tesla. The slow pace of travel, and plenty of space available to the truck tells me there adequate time to take control and avoid the collision. Fortunately a lesson learned at slow speed unlike others who recently lost their lives when using the system outside of its current capacity. Be thankful for the second chance.

I think the point is that this scenario should have EASILY been handled by AP1 or AP2 especially given how slowly things unfolded.

 

[kiwi]

https://www.navigantresearch.com/wp-assets/uploads/2018/01/LB-ADV-18-LB-pic.jpg
Navigant Research Leaderboard: Automated Driving Vehicles
The image says it all. The other companies are way ahead and while their designs look ugly, they actually seem to work.

Hey @AviP - the first URL seems dead, but the document in the second link says it all.

Its a fast moving field so there are interesting times ahead.
Most of the lead players require expensive lidar equipment (+ cameras), and so are focus on commercial
vehicles that can absorb the price (uber, taxis).
Tesla seems unique in going with only low resolution radar + cameras. Musk has stated that
he is confident that is enough.

Will Tesla solve AP with vision plus radar before the lidar requirements come down in price
and become built into cheaper cars? Time will tell.

 

[ErnieG]

Seems like radar didn't pick it up until it was mostly in your lane (which makes sense with shape and radar position), then FCW kicked on. May have been too slow for AEB?

Glad you are ok.

I was told the camera system that the autopilot uses does not identify the object as an obstacle until it is entirely in your lane, the ultrasonic sensor and radar do not at the present time (AP2.0) identify object but feed the processor and camera for decision making

Ernie

 

[Muzzman1]

Here's a similar cut-off, in better weather conditions in my AP1 MS, where the car began to enter my lane, my car veered and hit the brakes pretty hard to avoid the erratic driver.
I'm shocked the OP's car did not react at all!! Great info to keep in mind.
Thanks for taking one for the team OP!

 

[robby]

Hey @AviP - the first URL seems dead, but the document in the second link says it all.

Its a fast moving field so there are interesting times ahead.
Most of the lead players require expensive lidar equipment (+ cameras), and so are focus on commercial
vehicles that can absorb the price (uber, taxis).
Tesla seems unique in going with only low resolution radar + cameras. Musk has stated that
he is confident that is enough.

Will Tesla solve AP with vision plus radar before the lidar requirements come down in price
and become built into cheaper cars? Time will tell.

It isn't just that cameras are "enough." They are a much better and more reliable system. Lidar degrades heavily from atmospheric obstructions -- a SDV that is dependent on real-time lidar will not work anywhere in all weather conditions. Meanwhile, Nvidia has had a camera-based neural net that outperforms humans at object recognition in snow and fog dating back to 2016 (see their DRIVENet Demo on YoutTube).

The reason that so many companies are going the Lidar route is that it allows them to get a SDV on the road quickly. But it will be much harder for them to keep those vehicles on the road in bad conditions.

 

[mongo]

Here's a similar cut-off, in better weather conditions in my AP1 MS, where the car began to enter my lane, my car veered and hit the brakes pretty hard to avoid the erratic driver.
I'm shocked the OP's car did not react at all!! Great info to keep in mind.
Thanks for taking one for the team OP!

Hi,
Regarding the veer you mention, it appears the car's lateral behavior could be due to the missing yellow lines causing the car to realign to the lane.

 

[Peteski]

Medium. But if you watch the video with the audio on, you'll hear the collision warning go off about a blink before impact. I was fiddling with the GPS on my screen, and I recall hearing the beeping, slamming on the brake, but it was obviously too late to avoid hitting the truck.

I have my FCW set to "early" to give it the best chance of actually bleeping in time to take action. So far I've had very few false triggers, just the odd one while approaching stationary parked cars in narrow roads just before pulling out to pass them - which is fair enough and not even annoying at the time. Not sure "FCW Early" would have helped much in your case as the system seemed very late in even being aware of the truck, but I don't see any downside to setting FCW to early. It very rarely triggers in normal driving and when it does you probably need to be aware that you are getting close anyway.

 

[harcocjh]

Thanks for this post OP. I will have to check myself as to not get complacent with this system. I am taking delivery of my Model 3 in a few days and one of the primary reasons I have wanted a Tesla for so long is to have autopilot help my hour plus commute to and from work in mostly heavy Chicago traffic. Cars cut in all the time so this is really good to know ahead of time.

Hopefully over the next year or two issues like this one will be improved so that autopilot is better in situations like this.