First Tesla Autopilot Crash In China

[Tam]

It was announced in the TSLA stock thread.

Original thread in Chinese:

特斯拉自动驾驶中国首撞！无视障碍物加速_网易科技

I believe this is a Chinese version of the Switzerland scenario for colliding a stopped vehicle.

This is an Autopilot limitation that drivers should have learned from owner's manual as well as from Switzerland collision.

Here's the diagram of what happened:

The Model S (labeled blue) was locking on a moving leader car (labeled yellow).

The leader car swerved right to avoid colliding to a stopped vehicle (labled black).

The Model S (labeled blue) continued to drive straight to closely follow the leader car but also closely centering the lane without swerving and while it ignored the presence of the stopped vehicle.

The Model S driver now realized an impending collision and grabbed the wheel to swerve right but it was too late and it swiped the stopped vehicle.

Thankfully, the damage was minimal to both vehicles.

The swiping was so minor that the dashcam didn't visually display a sudden shock:

The moral of the story is: Please learn the limitations of the current Autopliot, be alert and avoid another replication of what already happened to others.

 

[bhzmark]

That's a tough problem to solve. AP has to decide whether to swerve out of lane and potentially side swipe a person in the next lane, or slam on the brakes to stop behind the parked car that is only partially blocking the lane.

That can be a tough problem for a human where the front car hides the stopped car, the front car has room to swerve in the next lane, but your following car is blocked in the next lane. btw a rock and a hard place.

a tough one. pointing out the need to remain alert and looking ahead.

 

[DSBS4]

On some LA freeways there's no shoulder in the fast/carpool lane. Some portions have a concrete sign support column that cuts into the lane a tiny bit.

Do you guys feel comfortable using AP on those portions?

 

[ecarfan]

This incident is being discussed in the Model S Driving Dynamics forum at Another driver doesn't take responsiblity...

 

[KZKZ]

You're right, very similar to the earlier Autopilot accident in Switzerland.

 

[stopcrazypp]

Same exact situation as the van crash. This is the situation every single TACC manual out there warns about. I believe everything that can be discussed was already discussed on that thread.

 

[Tam]

...I believe everything that can be discussed was already discussed on that thread.

So....

How will engineers have this scenario covered?

Would adding on a LIDAR help to avoid a crash with stationary vehicles?

Or is this just a software issue that requires a genius hacker to write the codes that work?

 

[cronosx]

I would say that it's a mix, the crash it's a software issue, you could simply stop the car or call the driver to act, this is well in realm of "only software update", for now they simply didn't address the problem for the false-positive it could create, they will surely address the problem in the future but it need more refinement.. said that it CAN'T be perfect because you don't have all the data you need to make the decision.

The problem is that of course stopping the car isn't the best choise, but enter another line to bypass the obstacle it's a thought decision to make if you don't have all the fact ( if you move too much on the other line this could means you crash on the right if it's not clear ), and the current sensor suite isn't good enought for this, the lateral/backward sonar has too much a limited range ( this is why you can't auto-change line without the driver ).
With 360° camera, this is a tought code to do, but it's doable. the car can say "ok the right lane is free, go a little on the right and it's ok" or say "the right line is full of car, it's better to stop"

But, in any case, you need to consider that this could bring at a crash in any case, the car in front of you maybe can make to the right lane, but maybe you can't.. and since you see the obsacle only after a certain time.. then maybe you don't have the needed time for the reaction, unless you stick to the "right distance", but i don't think you want to keep 100mt or similar from the car in front of you.. and of course maybe you can't stop a tesla in a certan distance, but can the car in your back? of course in this case it won't be your falt .. but again.. .. you see my point?

The right solution for this problem of course is car-to-car comunication, so what the car in front of you know, so do you and you have the right time for a reaction.

A lidar ( like in the google car ) is good only since it so up that it can see in front of the car in front of you ( and of course a camera can do the same ), but then.. if the car in front of you is a tir.. the google car lidar has 0 advantage since the sensor is not so high that it can see above the tir

 

[Tam]

On some LA freeways there's no shoulder in the fast/carpool lane. Some portions have a concrete sign support column that cuts into the lane a tiny bit.

Do you guys feel comfortable using AP on those portions?

Sounds like under-construction zone which I would not recommend Autopilot unless you are really super good at reacting at the last seconds to avoid accidents if Autopilot can't handle it.

 

[lklundin]

That's a tough problem to solve. AP has to decide whether to swerve out of lane and potentially side swipe a person in the next lane, or slam on the brakes to stop behind the parked car that is only partially blocking the lane.

That can be a tough problem for a human where the front car hides the stopped car, the front car has room to swerve in the next lane, but your following car is blocked in the next lane. btw a rock and a hard place.

a tough one. pointing out the need to remain alert and looking ahead.

Actually, I think this is a good example of a difficult situation that the AP has the potential to handle better than a human.

The AP can continuously calculate its optimal response to a number of potential scenarios, including a solid obstacle appearing and blocking the left part of the lane.

The AP would at all times know what other vehicles it has in its proximity (i.e. situational awareness), so at any given moment it would know how to adjust its speed (if needed) in order to safely change lanes to the right, and if that is not possible, then make sure to limit the speed so enough braking distance is available. (It hardly needs any reaction distance).

It looks like the current response could have been improved by shifting 5 cm (2 inches) more to the right, while signaling and perhaps using the horn to alert others to the danger.

 

[Tam]

...Actually, I think this is a good example of a difficult situation that the AP has the potential to handle better than a human...

My priority would be to focus on the reliability of automatically braking to avoid a frontal collision.

 

[lklundin]

My priority would be to focus on the reliability of automatically braking to avoid a frontal collision.

Fine, but braking is not always the best or even possible solution, so the AP needs to make a choice between steering and braking (or a combination).

Coming to a full stop because a lane is half-way blocked can cause that lane to be fully blocked (by you), in turn exposing you to the risk of a high-energy rear-ending with a likely whiplash injury. On a high-way with a heavy vehicle behind you this can kill all occupants of your vehicle.

 

[Tam]

Fine, but braking is not always the best or even possible solution, so the AP needs to make a choice between steering and braking (or a combination)...

Agreed. It is good to have more choices rather than just one single choice

 

[ecarfan]

The moral of the story is: Please learn the limitations of the current Autopliot, be alert and avoid another replication of what already happened to others

The moral of the story is RTFM.

 

[cronosx]

My priority would be to focus on the reliability of automatically braking to avoid a frontal collision.

You are in the field of software, it's not always so easy.
Just remember that doing from 0 to 90% of something in this field is at least 10 times easyer ( probabile 50-100 times ) to do the 90-100%, so just to put it in prospective, doing a full-autonomous from 0 to 90, it would take 1/8 of the time you need from doing the 90-95% of auto-breaking, and what you build for the full-autonomous can give you weapons for the auto braking.

Killing all the corner case isn't easy, and of course it seems to me that the auto-braking in this case was not at all needed afterall.. what damage the passenger had?

I would go full on the autonomy when we can be near the 90% of it, then i would go forward with the auto-brake, and then a little in the autonomy, then again autobreake and so on..

You don't really want the auto-breaking, you want a bullet-proof car that you can't crash in any way.. this is not the same thing as avoid death crash.. and this.. its the point for auto-emergency-brake

 

[Nikxice]

Actually, I think this is a good example of a difficult situation that the AP has the potential to handle better than a human.

The AP can continuously calculate its optimal response to a number of potential scenarios, including a solid obstacle appearing and blocking the left part of the lane.

The AP would at all times know what other vehicles it has in its proximity (i.e. situational awareness), so at any given moment it would know how to adjust its speed (if needed) in order to safely change lanes to the right, and if that is not possible, then make sure to limit the speed so enough braking distance is available. (It hardly needs any reaction distance).

It looks like the current response could have been improved by shifting 5 cm (2 inches) more to the right, while signaling and perhaps using the horn to alert others to the danger.

Your first sentence is spot on. Eventually AP will be faster and safer than a human while reacting to this type of lane blockage. AP has almost reached the point of what I refer to as, the horse won't ride you over the cliff mode. The assessment and response actions required for these scenarios will become automated. Even if the safest, yet less desirable option puts the AP vehicle stopping behind the van, AP could activate the hazard lights before a driver could reach the panel. Dangerous yes, but legally the stopped AP vehicle would have the right of way. Maintaining proper space management is the responsibility of those drivers following.

 

[GoTslaGo]

My main concern with this incident is Chinese regulators. I do not know Chinese laws, regulatory practices, etc... But given the recent "corruption" purges, taken at face value, then the bureaucracy is rigged. Some guy there yelling "it's AP's fault" and additional incidents like this could give these regulators in China an additional excuse to put the brakes on Tesla's potential growth in China. Also remember BYD is a big EV player in China, and a Chinese company.

While we see Tesla's growth in China to be a perfect match (great EVs for a market looking to promote EVs and reduce pollution), I suspect Chinese authorities may not be too keen on letting Tesla dominate the EV market vs. their own BYD (which incidentally is supported by Berkshire Hathaway and Bill Gates). There was actually an article which was linked by that Fukuta crazy short guy which implied that Chinese authorities were VERY ambivalent if not hostile toward adopting Tesla's technology because, well, it wasn't Chinese, and used Taiwanese parts (Fukuta). What he probably didn't realize is that it was the Taiwanese origin of the part that was the problem for China in that article.

Sigh, politics.

 

[Tam]

My main concern with this incident is Chinese regulators...

A few weeks after Tesla's announcement of the Autopilot Florida crash, China bans further runs of Autonomous Vehicles even though it was running at a frenzy pace prior to the prohibition.

China still has not banned an assisting system that still requires a driver at the helm such as Autopilot.

 

[Breezy]

'Self-driving' in spotlight again as China sees first Tesla autopilot crash

"We have never described autopilot as an autonomous technology or a 'self-driving car,' and any third-party descriptions to this effect are not accurate," the Tesla spokeswoman said.

And yet:

 

[Tam]

...And yet:....

The title from the local news station is misleading with the word "self-driving" but its contents are correct.

"I would characterize it as the ultimate cruise control," Neally explains about the autopilot.

Look ma! No hands! It's not an atonomous car, but it will stay in its lane, maintain speed, even brake to avoid crashes.

Neally says the car allows him to take his hands off the wheel for up to four minutes at a time. Then the car lets him know he needs to take the wheel at, least briefly, or else it looks for a spot to pull over and stop on the side of the road."

As with any newsheadline, it's a hook for you to get you to find out its content before you can see whether the headline is correct or not.