Will Tesla get to Level 5 Autonomy without LIDAR?

Will Tesla get to Level 5 Autonomy without LIDAR?

I've read a couple articles doubting whether Tesla Vehicles can reach level 5 autonomy without LIDAR. If this is true and Elon Musk has yet to come to this realization, then the current Models S, X and first generation 3 will all fall short of this goal. I personally don't want to have to buy another Tesla in a couple years just so I can have my car drive me home from work during a thunderstorm. | May 26, 2017

Unknown what will be needed for L5, so all wild speculation now. Most articles on this topic are in Seeking Alpha - where a very high percentage of articles are from people shorting Tesla stock. I wouldn't put much faith in them.

Keep in mind LIDAR has yet to prove itself, has it's own failings (does not work well in bad weather), unclear what the long term reliability is, if it can be made cost effective or how massive of computer is required to process the data. The software complexity could be dramatically higher than a vision based system as well, which means it will take longer to make work and could have many more bugs to kill.

Haggy | May 26, 2017

If Musk hasn't come to the realization yet, I suggest he start testing LIDAR. I'd suggest that he do it in a big way, such as by starting a company to launch rockets and test state of the art LIDAR to put him well ahead of the curve. I think he should do that for years, and then announce that he is going to have a self driving car once he has the proper experience. He hasn't already done that, has he?

Efontana | May 26, 2017

I think expectations for self driving are limited by the weather.If it can drive me home on most days, but asks for help when the bridge is flooded out, I am OK with that.

carlk | May 26, 2017

Elon think it's not necessary and I can not disagree with that. Do human need LIDAR to drive a car?

carlk | May 26, 2017

Haggy Actully a year or two ago some took a picture of a Tesla with LIDAR on top. The car also has a Stanford sign on it. So there could have been some work on that already.

johndoeeyed | May 26, 2017

Humans managed to get to Level 5 without LIDAR, so it is already proven possible.

hoffmannjames | May 26, 2017

Right now, LIDAR is very expensive which is probably a big reason why Tesla is not using it.

johndoeeyed | May 26, 2017

LIDAR is no longer very expensive due to solid state versions.
Tesla has also mentioned using RADAR in a pulsed mode like LIDAR.

Frank99 | May 26, 2017

>>>Humans managed to get to Level 5 without LIDAR, so it is already proven possible.
Yeah, but we'd never allow autonomous vehicles that caused 30,000 deaths a year in the US.

Mike83 | May 26, 2017

Multitasking is best left to AI, humans driving and texting and talking among other human errors increases the likely hood of accidents. AI will probably reduce accidents by 90%.

johndoeeyed | May 26, 2017

But you would allow them if they were shown to reduce that number to at least 3,000.

peter | May 27, 2017

Surely the key observation is that the quest for full autonomous driving is resulting in the analysis and categorisation of the almost infinite number of combinations of events that result in an accident. Current sensor and processing technology has now evolved to a point where a meaningful percentage of these combinations can now be managed but equally importantly the ones that can not be managed today can be better understood. These can be addressed by other means, either simply by improving road design or signage at specific black spots or improved sensor technology. Why not make it a priority for all manufacturers to be able to read road side "accident black spot" signs and automatically reduce speed.? We are in the transition phase from accident management being based on post accident statistical analysis to pre-accident active avoidance, we should build on the knowledge that we already have as we move forward to a technology limited ultimate solution. 30,000 avoidable deaths and countless more serious injuries is really unforgivable fro an intelligent species to tolerate.

dchuck | May 27, 2017

Yes, As others have pointed out AP2.0 cars have more sensors than any human being, so the problem is not one of Input but processing those inputs in real time.

The question is will they be able to do it with the current Nvidia hardware...

Remnant | May 28, 2017

@dchuck (May 27, 2017)

<< ... the problem is not one of Input but processing those inputs in real time. >>

More likely the need is for faster processing than that, but driver's intervention contingent on unexpected or unexpectable events is not likely to disappear, short of an ability to leap over obstacles or, simply to take to the air altogether.

It seems L5 is not likely as an earth-bound, or ground, capability, except on specially marked and communicative highways, for the foreseeable future.

johndoeeyed | May 28, 2017

You have no idea if there is a 'need is for faster processing than that". If you do have some actual knowledge in the area, then please share it with us.
The car only needs to be able to avoid an accident in those circumstances, not keep driving, just as a human should do.
I believe every major car manufacturer is working on L5 autonomy and believe it is coming 'real soon now', which is in the 'foreseeable future'.

SCCRENDO | May 28, 2017

Not sure whether this has been posted before but and interesting article.

SCCRENDO | May 28, 2017 | May 28, 2017

Elon is hedging a little. During the 1Q conference call, he suggested that the NVIDIA V100 processor may not be adequate to support L5. The good news is that the processor can be changed out via the glove box so a new car will not be needed in his opinion.

Great strides have been made in LIDAR development in recent years but neither Elon nor Amnon Sashua of Mobileye seem to think that it is necessary for L5. The case for cameras is strong. Sashua has claimed that three cameras are needed to duplicate human vision. Tesla is deploying eight of them. They cost about $5 each and have no wear out mechanism. Eight cameras and one radar transceiver currently cost about the same as the projected future cost of one solid state LIDAR system according to the following:

LIDAR can't penetrate fog or ssnow. Radar can.

For now Tesla appears to be on an achievable path. But I worry about the complexity of driving in a mixed environment of human and machine driven vehicles on poorly maintained roads often with temporary obstacles that can be confusing. If a human being has a collision while driving, it is no big deal. Happens all the time. But if a machine is driving, you can expect it to be a major news item. The publicity will be highly assymetric for a long time.

SCCRENDO | May 28, 2017

@george. I bow to your superior knowledge as I have just begun researching this topic. My gut impression is that what Elon does is a combination of great engineering progress with some marketing. Probably to get the perfect system would be too expensive. And I guess he is possibly aiming for Level 4 rather than Level 5. This may get him ahead of all the completion at an affordable level with Level 5 coming down the line.

johndoeeyed | May 28, 2017

L5 vehicles only need to be safer than human drivers.
They do not need to be perfect.

SCCRENDO | May 28, 2017

@Johndoe. Agreed. But will they be Level 4 or Level 5 without Lidar?

johndoeeyed | May 28, 2017

I have already stated my answer on this thread.

johndoeeyed | May 28, 2017

Elon has explicitly stated that they are going for no driver in 2 years i.e. L5

SCCRENDO | May 28, 2017

@Johndoe. I hope you are right. I'm keeping my Model S and getting my Model 3 purely for autonomous driving. The more autonomous for me the better.

johndoeeyed | May 28, 2017

It doesn't matter if I am right.
It matters if Elon is right. | May 28, 2017

I'm guessing that L5 operation > safety than human operation may not be acceptable for general release.

Might be that demonstrations of safety, whatever that means, of L5 operation may have to be several times better than human operation.

It will be interesting to see how L5 metrics re:safety are determined without exposing society to beta systems.
Maybe some sort of test track will be set up for comparison purposes.

AlMc | May 28, 2017

We are on AP 2.0 now with cameras, ultrasonics and radar. IMO (opinion) there is NO way we get to level 5 with the current AP 2.0 hardware/software. Redundancy with additional radar or Lidar will be needed to get there.

Personally, level 4 will be good enough for me.

as others have noted, EM was originally quite confident that AP 2.0 would get us to level 5. he has been backing off that recently.

johndoeeyed | May 28, 2017

Elon has stated that he expects it to be 10X better.
The software is being tested on existing vehicles right now by running in shadow mode.
It appears you know very little about Tesla's system.

AlMc | May 28, 2017

@George: Too prove level 5 is *safe* I think the 'beta testing' will be when using level 5 there will be feedback to the mothership that shows how level 5 *would have handled* a situation where the driver took over from level 4 in. certain situations.

IMO, we will all be beta testing for level 5 once level 4 becomes a reality.

AlMc | May 28, 2017

ERROR: 'When using level 4 we will be feeding back info to the mpothership' | May 29, 2017

@AlMc: Your points are well taken. We will be living in interesting times. I hadn't thought about redundancy. There isn't much of that in cars now. If a master cylinder fails, a car has a mechanical backup but if s steering linkage breaks, you are out of luck. Nonetheless, there could be a kind of double standard regarding redundancy. I worked in telecommunications systems for many years. Redundancy requirements were standard, especially for system controllers and transmission facilities.

Bernie: I was addressing your point: "L5 vehicles only need to be safer than human drivers." Not Elon's marketing hyperbole. I don't claim to know a lot about autonomous automobile systems. I doubt that you know much more. The OP is posing the point that LIDAR may be needed for L5. The rest of us are kicking the thought around.

I don't know about shadow mode but I do know people with AP2-equipped cars. Teslavision is still more of a promise than a fact. Until they get the basic functions like lane keeping, TACC, Auto-steer, and Auto lane changing working at least as well as the AP1 Mobileye stuff, I don't know what good shadow mode will do. They don't even have AEB working at highway speeds yet.

I'm not saying that Tesla isn't making progress. They are and someday they will get L2, 3, and 4 working and, by then, the answer to the OP's question and @AlMc's suspicions will probably be known.

Efontana | May 29, 2017

As for level 5, there is a lot of variation in human driving skills.

As for redundancy, it looks like Tesla is working to make the high definition map with lane discipline a redundant safe fail to the sensor controlled autonomy. Or something like that. The idea being you get redundancy not by exact duplication (n+n, or n+1 telecom approach), but by a belts and suspenders approach.

This should be more catastrophic failure robust, as the control paths are completely different - one event is unlikely to disable the control of the vehicle as the systems are not even geographically close to each other. So a sticky trash bag covering the camera pod does not require immediate human intervention. Rather the car goes into lane control based on the map, and collision control based on the radar, until the car is safe.

That will get you to level 5 compared to the 50th percentile human, without LIDAR.

(Note: I am not convinced that they do not use LIDAR to calibrate the object distance calculations eventually performed using images from the array of cameras. The model 3 with the roof rack and box may have had more to it. The fog light detail and holes on each box face, combined with what looks like precise leveling... There was an instrumentation look to that vehicle.)

Solutionist | May 29, 2017

One of the key things that you need to have full self driving is continuously updated mapping of roadways. Google currently just does not have the capacity and the model 3 I suspect is going to be that pathway for tesla to have the best roadway data.

Your AI Only goes so far you need to be able to track differences in what the car(s) see over time.

johndoeeyed | May 29, 2017

Elon was not using hyperbole. Tesla are continually testing their software in shadow mode, as well as their own internal tests, to gather the statistical data. 10X will be a statistically supported number, not pulled out of thin air.
The fact that you do not know about shadow mode points to you knowing very little about Tesla's autonomous driving system. | May 29, 2017

Very insightful, @Efontana. Roadway knowledge seems like a tricky thing to depend on. There are a lot of temporary lane closing shift for example for things like landscaping work, pothole repair, and shoulder work. There would have to be a closely regulated system put in place everywhere with databases continuously updated in order for this to work. Hard to see that happening.

One of the issues with two paths for control is the need for a vote that seems to eschew a false positive outcome. Maybe they will have three votes somehow with majority rule??? | May 29, 2017

"Lane closings" not "lane closing shift". iPad autocorrect | May 29, 2017

Bernie: I drive an AP1-equipped MX every day. That gives me some knowledge of Tesla's Autopilot features, especially the Mobileye version. It is hard to have knowledge of of Tesla's "autonomous driving system" as it doesn't yet exist. You neatly avoided mentioning my reference to your quote which was the point I was making. You said that L5 systems need only be safer than human drivers. They probably will need to be several times safer, don't you think?

As for Elon's hyperbole, he is a brilliant, driven person but he has repeatedly pronounced schedules and outcomes that have been absurdly optimistic. I waited 25 months for a car that was just around the corner. Eventually, he somehow manages to fulfill many of his promises but never on time. Hyperbole is exaggerated statements not meant to be taken literally. If the shoe fits...

Mike83 | May 29, 2017


johndoeeyed | May 29, 2017

Driving your MX gives you zero insight into how the autonomous system is validated to be safer than humans.
I did not 'avoid' anything. I have already told you that Elon has said they expect their driverless system to be 10X safer. I then explained how that would be validated. Why you would then ask "They probably will need to be several times safer, don't you think?" confounds me.
You then talk about hyperbole again. Please bother to read, and absorb, what I previously posted in response to your original statement. You appear quite ignorant of what Tesla is doing.

AlMc | May 29, 2017

If one takes everything that EM says as gospel, including his timelines, then the current AP 2.0 hardware is sufficient once the data collected in shadow mode and appropriate software is written then level 5 will be achieved.

Personally, I find it hard to believe which is my opinion. Until we see Tesla Vision reach 5.0 status all of us are giving are opinions. Some state their opinions as 'facts', which they are not.

AlMc | May 29, 2017

our opinions

johndoeeyed | May 29, 2017

NHTSA have said that AP1 is 40% safer than humans, from accident data.

johndoeeyed | May 29, 2017

Not opinion.

AlMc | May 29, 2017

When it gets to 100% ket me know.

johndoeeyed | May 29, 2017

It does not need to get to 100% i.e. it does not have to eliminate all possibility of an accident, and certainly cannot when humans are driving the other vehicles

SCCRENDO | May 29, 2017

As a cyclist I have some concerns
Tesla's self-driving tech is a danger to cyclists, robotics researcher says
Via #CNNMoneyStream

AlMc | May 29, 2017

Then let me know when ASP 2.0 hardware/software combo gets to level 5. It is a guess/opinion when/if that CAN happen.

You can frame it any way you like. Opinions vary on IF that can happen. Until it happens it is not a fact.

My opinion is that it can not. I accept that my opinion may be wrong.

johndoeeyed | May 29, 2017

That article is describing the current system, which is not intended to be used without driver supervision.
It is a different kettle of fish.

Remnant | May 29, 2017

@johndoeeyed (May 28, 2017)

<< @Remnant — You have no idea if there is a 'need is for faster processing than that". >>

Well, to implement Elon's vision of it, an L5 has to be better than the best human driver, in all respects and under all imaginable contingencies, including extra-terrestrial (such as Martian) ones.

<< The car only needs to be able to avoid an accident in those circumstances, not keep driving, just as a human should do. >>

It seems an L5 should be able to improvise driving, just as most humans would, even under conditions where no roads are available, or no road mapping is available or even possible, such as war, uncharted wilderness, or quasi-Apocalyptic environments — which I would call the heuristic mode.

Furthermore, an L5 should certainly include an amphibious or flood escape mode and even aerial and space or vacuum modes.

johndoeeyed | May 29, 2017

L5 does not need to be better than the best human driver. It only needs to be (significantly) better than the average human driver. Besides, that has nothing to do with whether the current hardware is sufficient. Nothing at all.

Humans should not keep driving many times that they do eg: when tired, in fog, when under the influence, when driving in flooding/ice, when just plain incompetent...