Autonomous Car Progress

[BlueBlazes]

True, except for Elon Musk. See anything Musky says that was, well, a crock of *sugar*, can be forgiven, ignored, etc. Like that there will be 1 million robotaxis on the road at the end of this year. Or that we will be able to sleep in our cars on the drive to work this year. Or that definitive statements are subsequently qualified with the words ‘feature complete’, known in politics as ‘double speak’ as it fools the less critical thinking.
See, so many obsequious acolytes have bound their self image to Musk, that resolving cognitive dissonance is already plugged into one’s daily planner. But, but, but on current trends cases/instances of such will soon be close to zero.
Additionally, I have it on good information that universities are soon closing worldwide & the peer review process is being ceased and we will simply await the latest Musk tweet for our education, since a Wolowitz, oops, I mean engineer is the fount of all knowledge

No question Elon is WAYYY overly optimistic with his timelines. But he never said there would be a million robotaxis by the end of this year. There will barely be a million Tesla's BUILT by the end of this year. He said that, regulators allowing, the service should be started by the end of this year. Don't think he'll make it in any case. He also didn't say L5 would be working by the end of this year. He said the fundamental problems would be solved by the end of the year and then it would be a matter of working on the edge cases. He didn't give a completion date for L5. He DID say earlier this year that there would be a version that would have a "greater-than-zero" chance of getting you from home to work without intervention. Based on that, if ANYONE makes it without intervention that becomes a true statement. So really not that bold an assertion.

 

[diplomat33]

Waymo's L4 is limited to a few geographic areas where it has mapped (and continuously re-maps to catch changes) every street, light, stop sign, traffic sign, etc., down to the millimeter. As I understand it, their LIDAR looks at the surroundings (pre-mapped, remember) and determines the exact location of the car. They use this to essentially drive like they're on rails. They don't need to read lane markers because they're always in the right place based on the maps. But even Waymo recently said they will still need safety drivers for at least a few more years. So that doesn't sound like a true L4 to me. If this isn't correct, I would appreciate it if someone would educate me.

This is not completely accurate. Yes, Waymo maps an area and uses lidar for precise localization. HD maps provide useful information to help increase safety:

"Before our cars drive in any location, our team builds our own detailed three-dimensional maps that highlight information such as road profiles, curbs and sidewalks, lane markers, crosswalks, traffic lights, stop signs, and other road features. Rather than rely on GPS, Waymo’s vehicles cross-reference their pre-built maps with realtime sensor data to precisely determine their location on the road."
Safety Report – Waymo

But that is just one piece of how they do autonomous driving. Waymo cars do not just follow a preset path based on the HD map "like on rails". Waymo cars use all their sensors to perceive the world and objects and figure out the best path and avoid hazards. Waymo cars are able to change paths based on sensor data to avoid hazards or when there is road construction.

Here is a diagram from a Waymo presentation that shows how their system works. You can see that they use both the Map and the Sensors (cameras, lidar, radar) for localization and also feed the Map and the Sensors into their Perception to detect all objects etc...

The Waymo Support page also has some information on how Waymo does autonomous driving:

How Waymo experiences the world
A combination of cameras, radar, microphones and LiDAR are the eyes and ears of every Waymo car.

Waymo uses…

• LiDAR to detect everything around the car, like pedestrians, cyclists, or other drivers
• Radar to help detect objects around the car and estimate their speed, like other motorists traveling around it
• Exterior cameras to see visual information, like whether a traffic light is red or green
• Microphones to hear audio cues, like the sound of police or emergency sirens

To help it plan a safe path ahead while driving, the car can see up to 3 football fields away in all directions.

How Waymo drives itself
Waymo software is the brain of the car. Through the software, the car makes sense of everything coming from the sensors, and uses the information in real time to drive itself.

Waymo cars can...

• Navigate on city streets to get from point A to B safely
• Avoid hazards including by slowing down, changing lanes, or braking completely for a stopped car
• Adjust to unexpected changes in the roads like road work or closed lanes
• Obey traffic laws like stopping at stop signs, and yielding to pedestrians or cyclists

Learn how Waymo drives - Waymo Help

This is an excellent video that explains how Waymo cars drive:

As you can see, it is much more than just mapping, localizing and following a path. IMO, it is L4 autonomy because the car is fully perceiving the world and is reacting in real-time to do all the driving tasks and drive without human intervention, just in a given area.

I think a big reason why Waymo geofences is because they don't want a customer who uses their ride-hailing service to select a destination that would take the car in areas that they have not fully validated yet.

Waymo has removed the safety driver for some drives where they are confident that it is safe to do so. But they are not ready to completely remove the safety driver for all rides yet.

 

[rdskill]

Biggest failing I see with cameras is they are useless when you need driver assist the most. Blinding heavy rain, fog, blinding sun in your eyes. I can see the message to "take control that AP is unavailable" just fine however.

Having look down to manually find the right wiper setting on a display instead of keeping your eyes on the road for a looming shadow of the car in front of you that you might rear-end is another problem.

 

[Nuclear Fusion]

No question Elon is WAYYY overly optimistic with his timelines. But he never said there would be a million robotaxis by the end of this year. There will barely be a million Tesla's BUILT by the end of this year. He said that, regulators allowing, the service should be started by the end of this year. Don't think he'll make it in any case. He also didn't say L5 would be working by the end of this year. He said the fundamental problems would be solved by the end of the year and then it would be a matter of working on the edge cases. He didn't give a completion date for L5. He DID say earlier this year that there would be a version that would have a "greater-than-zero" chance of getting you from home to work without intervention. Based on that, if ANYONE makes it without intervention that becomes a true statement. So really not that bold an assertion.

You need to research history before 2020

 

[Bob Denny]

Love my Model 3 but it is a little difficult to explain to non-Tesla owners why you paid $7,000 for FSD

I'll chime in here through everyone has already said it all. After spending a load of time and across 5 s/w updates, my 2020 MX with FSD isn't anything like full self driving, not even slightly. And that "amazing video of the AP" saving the guy from being smashed against the center wall.. well... there ain't no way I would ever take my car into "the slot" between a semi and a center wall, with traffic ahead, and right into the "reduced vision" (not blind) spot like that. My FSD takes me right along the right side of semis and into THAT bad spot. Very nerve wracking! Slowly passing a semi on the right is just plain stupid. There are so many aspects of defensive driving (yeah, like all those high-school defensive driving classes try to teach) that are not present in FSD. I think Elon might be hung out to dry by his techies. For reference see the Huemul Project and how Argentine dictator Peron got hung out to dry by a German scientist. No reference to @nuclearfusion above ha ha ha

 

[diplomat33]

From what I understand, Waymo's goal is to make their tech available to as many manufacturers as possible. Think of them like Microsoft Windows. Once the hardware can accommodate it, it'll work.
So if they get their way, they may end up grabbing the lion's share of the self driving market.

I think that is an apt comparison.

Waymo is aiming to provide FSD as a service:

"For the near future, Waymo will offer services through its proprietary fleet of autonomous vehicles. Long-term, however, Waymo is setting the stage to be a technology provider, not a truck manufacturer or fleet owner."
Waymo Outlines Plans to Integrate Autonomous Trucking in Fleets

 

[mikes_fsd]

From what I understand, Waymo's goal is to make their tech available to as many manufacturers as possible. Think of them like Microsoft Windows. Once the hardware can accommodate it, it'll work.
So if they get their way, they may end up grabbing the lion's share of the self driving market.

I think that is an apt comparison.

I disagree with the comparison, Microsoft did not dictate hardware, Waymo does not intend to just sell the software, they are packaging it as an integrated solution (roof puck sensor suite PLUS computer & their software).
And the current sensor suite Waymo has is expensive. (even the 5th generation).
So, I don't see them gaining market share, as this setup does not allow you (the car OEM) to find a place to optimize and reduce costs.

Just 1 or 2 weeks ago there was an article on a patent from Tesla where you could specify the sensor suite configuration for the NN's and then the FSD suite would know which set of NN's to run against the suite of sensors.

I would think you would need an approach closer to that.

Provide FSD computer with your software (Vision + Driving Policy) to the OEM, tell them ALL the types of sensors that your solution can handle (i.e. in personal computing terms which drivers you have available out of the box) and what minimums of sensors you will need to reach whatever SAE level. Let the OEM integrate the sensors into their vehicles as they see fit and plug into the FSD computer.

That seems like an approach capable of gaining market share.

 

[diplomat33]

I disagree with the comparison, Microsoft did not dictate hardware, Waymo does not intend to just sell the software, they are packaging it as an integrated solution (roof puck sensor suite PLUS computer & their software).
And the current sensor suite Waymo has is expensive. (even the 5th generation).
So, I don't see them gaining market share, as this setup does not allow you (the car OEM) to find a place to optimize and reduce costs.

I think part of the reason that Waymo is packaging the hardware and the software is because Waymo has developed custom in-house hardware with certain capabilities that make Waymo's FSD as good as it is. For example, Waymo brags about their cars being able to see 3 football fields in every direction with pinpoint accuracy. That's because of the superior cameras, lidar and radar that Waymo has developed in-house to achieve that. If you put Waymo's software on inferior hardware, you would get inferior FSD.

But the cost of the FSD hardware has already come down by a factor of 10. It will come down more. And the cost can't be that high if Waymo is putting twice the hardware on semi-trucks.

Just 1 or 2 weeks ago there was an article on a patent from Tesla where you could specify the sensor suite configuration for the NN's and then the FSD suite would know which set of NN's to run against the suite of sensors.

Source? That is news to me.

Provide FSD computer with your software (Vision + Driving Policy) to the OEM, tell them ALL the types of sensors that your solution can handle (i.e. in personal computing terms which drivers you have available out of the box) and what minimums of sensors you will need to reach whatever SAE level. Let the OEM integrate the sensors into their vehicles as they see fit and plug into the FSD computer.

That seems like an approach capable of gaining market share.

That sounds like the Mobileye approach.

But you are still thinking only from the point of view of FSD consumer cars. Remember that Waymo is aiming for ride-hailing robotaxis and trucking, not consumer cars, at this point.

 

[mikes_fsd]

Mobileye approach.

They do have Intel as their bank, so they have a good chance.

 

[mikes_fsd]

Source? That is news to me.

This is the one: Tesla's Neural Network adaptability to hardware highlighted in new patent application
Direct link to patent: US20200210832 SYSTEM AND METHOD FOR ADAPTING A NEURAL NETWORK MODEL ON A HARDWARE PLATFORM

In order to produce a concrete implementation of an abstract neural network, a number of implementation decisions about one or more of system’s data layout, numerical precision, algorithm selection, data padding, accelerator use, stride, and more may be made. These decisions may be made on a per-layer or per-tensor basis, so there can potentially be hundreds of decisions, or more, to make for a particular network. Embodiments of the invention take many factors into account before implementing the neural network because many configurations are not supported by underlying software or hardware platforms, and such configurations will result in an inoperable implementation.

 

[linux-works]

They do have Intel as their bank, so they have a good chance.

that's funny, to me. I can respect mobile eye quite a bit.

but, having worked for intel, personally, I have to say I have very little respect for that company, as they are at this point in time. inertia carries them in many ways, but intel is not an innovator anymore. they can still buy companies, but intel is mostly a shell of the greatness they once were. shame, really.

 

[mikes_fsd]

that's funny, to me. I can respect mobile eye quite a bit.

but, having worked for intel, personally, I have to say I have very little respect for that company, as they are at this point in time. inertia carries them in many ways, but intel is not an innovator anymore. they can still buy companies, but intel is mostly a shell of the greatness they once were. shame, really.

That is why I said they can be used as a bank not as the innovator. Their money is still good...

 

[Tam]

It's interesting that Navistar has partnered with new startup TuSimple instead of proven Waymo for L4 trucks with the goal date of 2024:

Navistar to Produce Self-Driving Trucks by 2024

Its sensor range is 1km in L4 instead of the average range of 250m for L2 (football field length is 91.44m or 300feet):

Its processing power is 400 trillion operations per second.

The assignment for TuSimple is to reduce the complexity of unfamiliar environments to known and prepared fixed routes and depots which is simpler than ride-hailing service that has way more many destinations to work on.

 

[linux-works]

That is why I said they can be used as a bank not as the innovator. Their money is still good...

agreed.

if intel offers money, take it

 

[diplomat33]

This is the one: Tesla's Neural Network adaptability to hardware highlighted in new patent application
Direct link to patent: US20200210832 SYSTEM AND METHOD FOR ADAPTING A NEURAL NETWORK MODEL ON A HARDWARE PLATFORM

Thanks

 

[john5520]

I disagree with the comparison, Microsoft did not dictate hardware, Waymo does not intend to just sell the software, they are packaging it as an integrated solution (roof puck sensor suite PLUS computer & their software).

Actually Microsoft did and still works closely with OEMs to ensure Windows compatibility and reliability on their systems. So, they do have some influence over the hardware and vice versa. Waymo also works with OEMs to help ensure their system will work as expected.

And if Tesla is thinking about also being a technology provider, they will surely have to work with OEMs to ensure their systems work as intended as well. Sensors and cameras would have to be placed in certain areas to be most effective for example.

Also, regarding Tesla's approach, the FSD computer IS hardware. And if they're going to dictate the type of sensors that can be used, they'll still be offering a software / hardware solution that's somewhat limited. And whether it's actually going to be cheaper is questionable. What will they charge to cover their development costs? Developing and constantly tweaking the algorithms for the NN isn't cheap. Then you have reliability. What's the point in having a cheaper solution if it's going to be less reliable? And the answer to whether the NN takes a lot of time to develop is evident in our cars.

And the current sensor suite Waymo has is expensive. (even the 5th generation).
So, I don't see them gaining market share, as this setup does not allow you (the car OEM) to find a place to optimize and reduce costs.

Well that's assuming the price of the hardware remains the same over time, which is almost never the case with hardware / technology in general. Economies of scale and improvements in engineering help to lower costs. Also, what they're testing now as mentioned already, is aimed at commercial enterprises. What will actually show up on consumer cars is most likely going to be less bulky and less expensive.

Just 1 or 2 weeks ago there was an article on a patent from Tesla where you could specify the sensor suite configuration for the NN's and then the FSD suite would know which set of NN's to run against the suite of sensors.

I would think you would need an approach closer to that.

Provide FSD computer with your software (Vision + Driving Policy) to the OEM, tell them ALL the types of sensors that your solution can handle (i.e. in personal computing terms which drivers you have available out of the box) and what minimums of sensors you will need to reach whatever SAE level. Let the OEM integrate the sensors into their vehicles as they see fit and plug into the FSD computer.
That seems like an approach capable of gaining market share.

Sure, I believe the NN will develop into something akin to a thinking human, but when? There's obviously a lot of work left to be done. All the while, mapping technology and lidar systems will be improving to meet the demands of autonomous vehicles. The more I research the different approaches, the more I believe the Lidar / radar and HD maps approach is the
superior one for at least the near term.

It's not that I'm counting Tesla out. They have so much more to offer in their cars than just automation. It's just that their approach, IMO, will take a longer time to come to fruition. And by fruition, I mean level 4-like automation. And like I've said before, I didn't buy a Tesla expecting L4 or 5 FSD anytime soon. When they work out the kinks with what they have now, I'll be content.

What I foresee for the future though, is the melding of both technologies into one system. Just image an EV with lidar + radar + HD maps + cameras and the NN?

I think part of the reason that Waymo is packaging the hardware and the software is because Waymo has developed custom in-house hardware with certain capabilities that make Waymo's FSD as good as it is. For example, Waymo brags about their cars being able to see 3 football fields in every direction with pinpoint accuracy. That's because of the superior cameras, lidar and radar that Waymo has developed in-house to achieve that. If you put Waymo's software on inferior hardware, you would get inferior FSD.

But the cost of the FSD hardware has already come down by a factor of 10. It will come down more. And the cost can't be that high if Waymo is putting twice the hardware on semi-trucks.

Agreed. Couldn't have said it better.

 

[diplomat33]

No question Elon is WAYYY overly optimistic with his timelines. But he never said there would be a million robotaxis by the end of this year. There will barely be a million Tesla's BUILT by the end of this year. He said that, regulators allowing, the service should be started by the end of this year. Don't think he'll make it in any case. He also didn't say L5 would be working by the end of this year. He said the fundamental problems would be solved by the end of the year and then it would be a matter of working on the edge cases. He didn't give a completion date for L5. He DID say earlier this year that there would be a version that would have a "greater-than-zero" chance of getting you from home to work without intervention. Based on that, if ANYONE makes it without intervention that becomes a true statement. So really not that bold an assertion.

I think the "1 million robotaxis" part might be a case of shoddy journalism.

If you look at news article like this one, they quote Elon as saying this during Autonomy Day:

"I feel very confident predicting 1 million autonomous robo-taxis for Tesla next year," Musk told a room of investors and Wall Street analysts at the company's Palo Alto, California, headquarters. "Not in all jurisdictions because we won't have regulatory approval everywhere, but I'm confident we will have regulatory approval at least somewhere, literally next year. "
Elon Musk says Tesla will have 1 million robo-taxis on the road next year, and some people think the claim is so unrealistic that he's being compared to PT Barnum

But if you watch Autonomy Day, Elon actually says this:
"I feel very confident predicting autonomous robo-taxis for Tesla next year. Not in all jurisdictions because we won't have regulatory approval everywhere, but I'm confident we will have regulatory approval at least somewhere, literally next year. "

At 3:06:13 mark:

Same quote, but no "1 millon robotaxis".

Unless, Elon mentioned "1 million robotaxis" somewhere else that I am missing? But it certainly appears like the reporters took a quote from Elon and inserted "1 million robotaxis" where Elon did not actually say it.

 

[bd7349]

I think the "1 million robotaxis" part might be a case of shoddy journalism.

If you look at news article like this one, they quote Elon as saying this during Autonomy Day:

"I feel very confident predicting 1 million autonomous robo-taxis for Tesla next year," Musk told a room of investors and Wall Street analysts at the company's Palo Alto, California, headquarters. "Not in all jurisdictions because we won't have regulatory approval everywhere, but I'm confident we will have regulatory approval at least somewhere, literally next year. "
Elon Musk says Tesla will have 1 million robo-taxis on the road next year, and some people think the claim is so unrealistic that he's being compared to PT Barnum

But if you watch Autonomy Day, Elon actually says this:
"I feel very confident predicting autonomous robo-taxis for Tesla next year. Not in all jurisdictions because we won't have regulatory approval everywhere, but I'm confident we will have regulatory approval at least somewhere, literally next year. "

At 3:06:13 mark:

Same quote, but no "1 millon robotaxis".

Unless, Elon mentioned "1 million robotaxis" somewhere else that I am missing? But it certainly appears like the reporters took a quote from Elon and inserted "1 million robotaxis" where Elon did not actually say it.

I noticed that starting at 2:45:40 Elon says the following:

“There’s three steps to self driving you know, there’s being feature complete, then there’s being feature complete to the degree where we think the person in the car does not need to pay attention, and then at a reliability level where we are also able to convince regulators that that is true. So there’s kind of like three levels — we expect to be feature complete in self driving this year (2019) and we expect to be confident enough from our standpoint to say that we think people do not need to touch the wheel, look out of the the window, sometime probably around I dunno, second quarter of [2020]. And then we expect to get regulatory approval, at least in some jurisdictions, for that towards the end of [2020] That’s roughly the timeline that I expect things to go on.”

Now, they didn’t achieve these goals exactly as stated but he did say that this was “roughly” the timeline that he expects. The reason I find this interesting is that it partially matches up with Elon’s statements in the most recent earnings call in Q1 where he also said they still expect to roll out FSD by the end of the year (exact quote: “extremely confident that it'll be possible to do a drive from your home to your office most of the time with no interventions by the end of the year”), and how he still expects they’ll go for regulatory approval towards the end of the year/early next year.

That, combined with Elon’s recent tweet about “a lot of new functionality” coming in the next 2-4 months when they roll out the new Autopilot that they’ve been working on, makes me think Elon’s timelines on FSD are converging for once. In previous times they’ve just gotten further and further pushed out, but that doesn’t seem to be happening now.

Thoughts?

 

[diplomat33]

I noticed that starting at 2:45:40 Elon says the following:

“There’s three steps to self driving you know, there’s being feature complete, then there’s being feature complete to the degree where we think the person in the car does not need to pay attention, and then at a reliability level where we are also able to convince regulators that that is true. So there’s kind of like three levels — we expect to be feature complete in self driving this year (2019) and we expect to be confident enough from our standpoint to say that we think people do not need to touch the wheel, look out of the the window, sometime probably around I dunno, second quarter of [2020]. And then we expect to get regulatory approval, at least in some jurisdictions, for that towards the end of [2020] That’s roughly the timeline that I expect things to go on.”

Now, they didn’t achieve these goals exactly as stated but he did say that this was “roughly” the timeline that he expects. The reason I find this interesting is that it partially matches up with Elon’s statements in the most recent earnings call in Q1 where he also said they still expect to roll out FSD by the end of the year (exact quote: “extremely confident that it'll be possible to do a drive from your home to your office most of the time with no interventions by the end of the year”), and how he still expects they’ll go for regulatory approval towards the end of the year/early next year.

That, combined with Elon’s recent tweet about “a lot of new functionality” coming in the next 2-4 months when they roll out the new Autopilot that they’ve been working on, makes me think Elon’s timelines on FSD are converging for once. In previous times they’ve just gotten further and further pushed out, but that doesn’t seem to be happening now.

Thoughts?

Elon's original timeline looked liked this:

End 2019: Feature Complete
Q2 2020: "hands-free"
End 2020: Regulators approve FSD in some jurisdictions

Remember that we are not at Feature Complete yet. If Tesla delivers the new functionality in 2-4 months with the rewrite, I doubt that we would jump straight to Elon's second level of going hands-free. Optimistically, the rewrite would just get us to Feature Complete, ie Elon's first level. So we would just be at Elon's first level by end of 2020 instead of 2019. So Feature Complete got pushed back at least 1 year. I am not seeing any convergence.

I would also point out that Elon's third level of regulators approving FSD only in some jurisdictions would be SAE L4 autonomy. So it would still fall short of Elon's L5 autonomy goal.

 

[Mardak]

Unless, Elon mentioned "1 million robotaxis" somewhere else that I am missing?

Just needed to go a little bit further in the video at 3:13:50: "From our standpoint, if you fast forward a year -- maybe a year and 3 months (July 2020), but next year for sure, we will have over a million robo-taxis on the road. The fleet wakes up with an over-the-air update. That's all it takes."

Where it looks like he's using the term "robo-taxi" here as any Tesla with hardware and software capable of driving without a driver -- not necessarily all owners are participating in the robo-taxi fleet nor that any regulator has approved it for driving without a driver. This is consistent with his original prediction of "hands free" FSD.