New article: "New Research May Vindicate Tesla's Choice To Forego LIDAR"

[strangecosmos]

My new article: New Research May Vindicate Tesla's Choice To Forego LIDAR.

Summary:

• New research finds that a self-driving car can use cameras to determine distances to within 10 centimeters (3.9 inches) of accuracy.
• 10 cm of accuracy is most likely sufficient for self driving, making LIDAR unnecessary.
• The accuracy of cameras was only tested at low speeds, but the evidence points toward accuracy being good enough for high speeds.
• Foregoing LIDAR may be a strategic master stroke, giving Tesla a multi-year lead in self-driving over all competitors.

I'm looking for critical feedback on the technical details of this article from people who know more than me about photography, LIDAR, computer vision, machine learning, and robotics. I have linked to the Google AMP version of the article, meaning I don't get any money when you click. I'm sharing this article here purely for the purposes of learning.

Please keep general comments about self-driving car safety, redundancy, and ethics to other threads that already exist for those topics. I want to focus on the technical details in this thread. Thank you!

 

[Pale_Rider]

Why are you posting this again? Your original post about this is just a few lines down?

 

[Tiger]

A sensor is just a plugin in a autonomous driver solution. It does not give Tesla any lead if they use plugin x instead of plugin y for measuring surroundings. The measurements of the plugin are abstracted and the driving logic does not even know what the physical devices are, and the driving logic can even be running in a simulated environment with simulated measurements.

 

[strangecosmos]

Why are you posting this again? Your original post about this is just a few lines down?

This article weaves together a lot of information that is not in that thread.

 

[Tam]

...New research...

There are many new researches that promises the moon for a near future but this is still a research and not a practical proof that we don't need LIDAR.

Tesla says we it doesn't need LIDAR but there's still no proof that its safety matches Google's LIDAR system just yet!

 

[strangecosmos]

A sensor is just a plugin in a autonomous driver solution. It does not give Tesla any lead if they use plugin x instead of plugin y for measuring surroundings.

The advantage is time. Using cheaper sensors enables deployment sooner.

 

[Tiger]

The advantage is time. Using cheaper sensors enables deployment sooner.

There isn't any such advantage, anybody can buy any sensors at any time and plug them into their system and update respective dynamic parameters so that the algorithm can adapt to new sensor. It does not matter what sensors (real or simulated) you developed your driving algorithm with (or if it does, then you have a very bad spaghetti algorithm design because sensor technology develops very fast and it [driving algorithm] should allow for pluggable sensors [how many and what kind of each]).

 

[strangecosmos]

There isn't any such advantage, anybody can buy any sensors at any time and plug them into their system and update respective dynamic parameters so that the algorithm can adapt to new sensor.

By deploying cars with self-driving hardware in late 2016, Tesla is able to collect a lot of driving data, especially going into Model 3 production. Competitors aren't planning to deploy cars with self-driving hardware until around 2021. Part of this is the cost of LIDAR (it's expensive to retrofit cars with sensors after they've been delivered to customers). That means Tesla may have enough driving data to develop and validate self-driving software and deploy it in production cars before any competitor even makes a production car with self-driving hardware.

 

[strangecosmos]

Let me put it very simply: if everyone but Tesla is using LIDAR, and LIDAR won't be affordable until 2027, then Tesla will have self-driving cars before everyone else.

 

[Tiger]

Let me put it very simply: if everyone but Tesla is using LIDAR, and LIDAR won't be affordable until 2027, then Tesla will have self-driving cars before everyone else.

Not necessarily. Nissan already has the solution, they just need to compress the technology, which we know is more predictable than algorithm development in general:

Furthermore, don't forget that a driving algorithm is software, parts of it or whole can be copied any time and there goes your any-kind-of algorithmical advantage.

 

[setoKaiba]

Let me put it very simply: if everyone but Tesla is using LIDAR, and LIDAR won't be affordable until 2027, then Tesla will have self-driving cars before everyone else.

Lidar for self driving car is already affordable today.

The 2018 Audi A8, the first self driving Level 3 car according to the SAE has a front facing Lidar with 140 degrees field of view and works up to 200 meters (656.168 ft).

Hands on -- and off -- with the Audi A8's Level 3 self-driving system

This is the same EXACT lidar that is used on Audi's A7 Jack Highway High speed test cars and can drive up to 80mph on divided highways and change lanes independently.

Audi Level 4 cars being released in 2020 and tested on the elaine platform will use the EXACT same lidar only with the addition of another one (making it two).

This defeats the notion that Lidar won't be affordable until 2027 as the same lidar are in full production today in 2017.

Almost all automotive company which are planning to deploy Level 4 Highway capable consumer cars have their lidar solution all figured out.

- BMW 40 self-driving prototype vehicles is using a similar Lidar as Audi. They have two platform, "one with Level 3 in mind and the other meant to get the automaker ready for Levels 4 and 5."... "The company is using three cameras behind the windshield, of varying focal length, which look like any other standard equipment. The LiDAR sensors (one on the Level 3 car, three on the Level 4/5 cars)" If you look at the pictures of the lidar on their car, it looks almost identical to the Audi Lidar, I won't be surprised if they use the same Lidar.

Now I can mention other companies like Nissan who plans to release a high speed highway L3 self driving car in japan next year 2018 and they are using up to 4 Lidars. Then others like Volvo who also plan to release a L4 Highway car in 2021 and uses one forward facing Lidar. Then you have tons of other companies like Hyundai who uses IBEO Lidar and many more.

So with the fact that you have Lidar already in production and other companies using the same identical or similar Lidar for release in similar timeline and then multiple new solid state lidar being projected to be ready for full production in 2018/2019 for $100-$200 (Innoviz, LeddarTech, Quanergy, Strobe, many more).

The statement that "LIDAR won't be affordable until 2027" is patently, demonstratively and factually false!
Your analysis completely ignores everything.

 

[AnxietyRanger]

Your analysis completely ignores everything.

I don't doubt @Trent Eady is sincere, I really don't, but this is the problem when people get fixated on some thought. It is like the old saying, for a person with a hammer, everything looks like a nail.

IMO some times people are too fixated on the vision story sold by Elon Musk to really evaluate Lidar fairly. It seems more like he is seeking to disprove Lidar, more than to understand it.

 

[Tam]

...LIDAR won't be affordable until 2027...

In 2012, Model S battery pack was about $40,000. Today, 5 years later, it's about $12,000.

Velodyne mechanical-spinning LIDAR used to cost $80,000. As of last year, it costed about $8,000.

The company says it will be able to sell you $50 each for its non-spinning solid state LIDAR in high quantities.

What's the doom and gloom about LIDAR delays?

 

[lunitiks]

You got a little hammer in your pocket too, don't you @AnxietyRanger? I sure do sometimes. Glass house and all that

 

[AnxietyRanger]

You got a little hammer in your pocket too, don't you @AnxietyRanger? I sure do sometimes. Glass house and all that

Hammer? Pfft.

I'm throwing stones!

 

[strangecosmos]

The statement that "LIDAR won't be affordable until 2027" is patently, demonstratively and factually false!
Your analysis completely ignores everything.

I certainly don't think 2027 is actually a realistic year. I was using an extreme hypothetical to make the point that time is not an irrelevant factor.

The thing to keep in mind about low cost lidar units is how their specs compare to the lidar units used by Waymo and others: Cheap lidar sensors are going to keep self-driving cars in the slow lane You can get a cheap, low-quality lidar unit, but it isn't capable of the stuff that high-quality units like Waymo's are like object recognition or automatic emergency breaking at highway speeds. At that point, cameras, radar, and ultrasonics are pretty much doing all the work anyway.

Audi is planning to use lidar for traffic-aware cruise control (TACC) on divided highways at a max of 37 mph, but Tesla already does TACC at higher speeds without lidar. The Level 3 classification is really about the level of safety risk a car company is willing to take, not the intrinsic technological capabilities of a system. Tesla could probably make its TACC Level 3 at under 38 mph as well.

Talking about Level 4 highway driving shows the fuzziness of the SAE definitions. Just automating highway driving is not what most people have in mind when you say "self-driving".

 

[strangecosmos]

The company says it will be able to sell you $50 each for its non-spinning solid state LIDAR in high quantities.

What's the doom and gloom about LIDAR delays?

No timeline on this except to say 1 million units/year is planned for the early 2020s. Tesla deployed in 2016. That's a delay of ~4-8 years. Huge from a financial and safety perspective.

 

[croman]

Audi is planning to use lidar for traffic-aware cruise control (TACC) on divided highways at a max of 37 mph, but Tesla already does TACC at higher speeds without lidar. The Level 3 classification is really about the level of safety risk a car company is willing to take, not the intrinsic technological capabilities of a system. Tesla could probably make its TACC Level 3 at under 38 mph as well.

Talking about Level 4 highway driving shows the fuzziness of the SAE definitions. Just automating highway driving is not what most people have in mind when you say "self-driving".

The SAE definitions are fairly clear. They might be subject to revision but I do believe there is a technological superiority to a L3 or L4 vs. the best L2. It all boils down to capability and confidence. Could Tesla call its highway AP L3? Sure, soon, but Audi's system is L3 because the system itself does what it needs without any user input. That is close to what EAP is marketed as doing but clearly EAP is not real right now. The only thing real about EAP is the fact its trying to reach parity with AP1 and then, perhaps, we'll start seeing Tesla's system pushed to match L3 systems (but without Audi's silly speed restrictions).

The risk and whatnot a manufacturer bears is really only applicable in Germany. In the US, product liability litigation might look to what the system was responsible for vs. the driver but I think Tesla's lawyers will put language into any EAP/FSD feature that gives Tesla an out in terms of liability. All products rely on warnings, statements of intended use, and user level restrictions to ensure that liability is limited as much as possible. Tesla will be no exception.

 

[setoKaiba]

I certainly don't think 2027 is actually a realistic year. I was using an extreme hypothetical to make the point that time is not an irrelevant factor.

It is irrelevant because the time factor has already been beat. Lidars used in L4 Highway Self driving cars ARE ALREADY in production TODAY!

And that's with 140 degrees field of view and range of up to 200 meters (656.168 ft).

The thing to keep in mind about low cost lidar units is how their specs compare to the lidar units used by Waymo and others: Cheap lidar sensors are going to keep self-driving cars in the slow lane You can get a cheap, low-quality lidar unit, but it isn't capable of the stuff that high-quality units like Waymo's are like object recognition or automatic emergency breaking at highway speeds. At that point, cameras, radar, and ultrasonics are pretty much doing all the work anyway.

All lidar being deployed for L3/L4 highway for 2020/2021 currently uses a Lidar with small number of lasers compared to the more High Resolution Lidars. Unlike what you said, they are used for lowly things such as AEB but are also used for object recognition and classification.

Specifically the lidar that Audi & BMW uses for their L3/L4 high speed highway system are:

"Valeo SCALA® scans the area in front of the vehicle and detects vehicles, motorbikes, pedestrians and static obstacles. Using the collected data, the scanner enables the creation of a map of the environment allowing to analyze and anticipate events around the vehicle.
This technology as well serves to enhance active safety by initiating measures such like evasion maneuvers or emergency braking."

Even the lowest resolution automotive lidar are more high quality/higher resolution than radar and ultrasonics as you can see below from a mobileye presentation:

Imgur: The most awesome images on the Internet

In a L4 Highway system, you don't need a lidar with 16, 32 or 64 lasers collecting millions of points per second because you are not operating in a dense urban environment.

Audi is planning to use lidar for traffic-aware cruise control (TACC) on divided highways at a max of 37 mph, but Tesla already does TACC at higher speeds without lidar.

Audi L3 system is not TACC. Its a traffic jam pilot and the first of its kind.

The Level 3 classification is really about the level of safety risk a car company is willing to take, not the intrinsic technological capabilities of a system.

Level 3 IS based on the intrinsic technological capabilities of the system as that leads to their confidence of the automaker allowing the driver to NOT pay attention.

Tesla could probably make its TACC Level 3 at under 38 mph as well.

Tesla AP already has encountered hundreds of reported accident, many occurring under 37 mph. So no Tesla can't make TACC Level 3.

Talking about Level 4 highway driving shows the fuzziness of the SAE definitions. Just automating highway driving is not what most people have in mind when you say "self-driving".

This shows you're not well versed in the self driving industry. As There's nothing fuzzy about Level 4 as defined by SAE and that is what All automotive company are aiming towards for 2020/2021 consumer releases. Likewise all ride-hailing program are also aiming for Level 4 autonomy.

The SAE is used by the NHTSA and the U.S Department of Transportation. In addition all self driving legislature from EVERY state is based solely on the SAE definition, including the one that just passed the House and the Senate Committee and will be signed by President Trump into law sometime next year once it is voted by the full house and senate. So if you are going to continue discussing SDC tech, I suggest you get knowledgeable in it.

 

[calisnow]

It is irrelevant because the time factor has already been beat. Lidars used in L4 Highway Self driving cars ARE ALREADY in production TODAY!

And that's with 140 degrees field of view and range of up to 200 meters (656.168 ft).



All lidar being deployed for L3/L4 highway for 2020/2021 currently uses a Lidar with small number of lasers compared to the more High Resolution Lidars. Unlike what you said, they are used for lowly things such as AEB but are also used for object recognition and classification.

Specifically the lidar that Audi & BMW uses for their L3/L4 high speed highway system are:

"Valeo SCALA® scans the area in front of the vehicle and detects vehicles, motorbikes, pedestrians and static obstacles. Using the collected data, the scanner enables the creation of a map of the environment allowing to analyze and anticipate events around the vehicle.
This technology as well serves to enhance active safety by initiating measures such like evasion maneuvers or emergency braking."

Even the lowest resolution automotive lidar are more high quality/higher resolution than radar and ultrasonics as you can see below from a mobileye presentation:

Imgur: The most awesome images on the Internet

In a L4 Highway system, you don't need a lidar with 16, 32 or 64 lasers collecting millions of points per second because you are not operating in a dense urban environment.



Audi L3 system is not TACC. Its a traffic jam pilot and the first of its kind.



Level 3 IS based on the intrinsic technological capabilities of the system as that leads to their confidence of the automaker allowing the driver to NOT pay attention.



Tesla AP already has encountered hundreds of reported accident, many occurring under 37 mph. So no Tesla can't make TACC Level 3.




This shows you're not well versed in the self driving industry. As There's nothing fuzzy about Level 4 as defined by SAE and that is what All automotive company are aiming towards for 2020/2021 consumer releases. Likewise all ride-hailing program are also aiming for Level 4 autonomy.

The SAE is used by the NHTSA and the U.S Department of Transportation. In addition all self driving legislature from EVERY state is based solely on the SAE definition, including the one that just passed the House and the Senate Committee and will be signed by President Trump into law sometime next year once it is voted by the full house and senate. So if you are going to continue discussing SDC tech, I suggest you get knowledgeable in it.

Welcome back bladerskb!