What GPS receiver do HW2/AP2 cars use?

[strangecosmos]

Thank you @lunitiks and @kdday for confirming!

@BinaryField, how much does the math described in the paper help the accuracy? For example, if you’re 450 km away from a base station and your accuracy is down to 4 metres, would you able to use the math to get this down to 40 cm? How should I think about that?

 

[MiaGiada]

Even if the GPS could be under 1cm accurate, how does it account for obstacles and detours? It doesn't. You need to solve real-time vision.

 

[strangecosmos]

Even if the GPS could be under 1cm accurate, how does it account for obstacles and detours? It doesn't. You need to solve real-time vision.

Yes, I agree. I don’t think self-driving cars should operate in zero visibility conditions. I just thought it was an interesting thought experiment.

1 cm GPS would provide redundancy for HD maps and real time visual SLAM. But it looks like Tesla’s GPS is only accurate to 1.5 m, unless Tesla starts installing is own GPS base stations at every Supercharger, store, and service centre.

 

[strangecosmos]

Wait! I think this could work... in AUSTRALIA! The country is blanketed with GPS base stations. Almost everywhere you would drive in Victoria is within 30 km of a base station. Take a look at this map! Incredible.

SmartNet Aus Site Map | SmartnetAus

2-centimetre differential GPS would totally work in Victoria. Very very very cool. HW2 Teslas have an awesome source of redundancy for the multi-camera system. A car can place itself within the HD maps using cameras and its rough (1.5 m) GPS location, but with 2-centimetre GPS it can pinpoint its exact location as accurately as lidar. Amazingly cool.

Compare the base station map to the population map. Almost all of the populous areas are covered by base stations, and the more densely populated areas are more densely covered.

 

[BinaryField]

Thank you @lunitiks and @kdday for confirming!

@BinaryField, how much does the math described in the paper help the accuracy? For example, if you’re 450 km away from a base station and your accuracy is down to 4 metres, would you able to use the math to get this down to 40 cm? How should I think about that?

There's a lot of background here required to understand what the paper is trying to do. Basically, there are two measurements you can get out of the GPS signals to determine your position: pseudorange and carrier phase. Pseudorange tends to be very noisy and blunt, but it gives you an unambiguous ranging measurement. On the other hand, carrier phase is very clean and precise, but it has an ambiguous term more or less equal to the number of full carrier wavelengths between the receiver and GPS satellite. Most common applications (e.g. your phone) use just the pseudorange measurements. For high accuracy applications, it is desirable to process the carrier phase measurements, but it takes a lot of work to resolve the ambiguity in the number of wavelengths. The process that tries to solve for this number is called integer ambiguity fixing / resolution.

The paper is about finding a more computationally efficient way to take advantage of the precise nature of carrier phase measurements. It doesn't have anything to do with improving accuracy over long baselines. The accuracy drop-off over long baselines, as mentioned before, is primarily driven by non-cancelling atmospheric effects between base station and rover. The only way to get rid of these effects is to have highly accurate atmospheric models, or to try to cancel them out by processing DGPS measurements in a special way. Both of these methods have a lot of problems: the former is just hard, if not impossible to get in real time; the latter introduces a high amount of noise into the measurements, which most likely isn't good enough for high accuracy applications.

Wait! I think this could work... in AUSTRALIA! The country is blanketed with GPS base stations. Almost everywhere you would drive in Victoria is within 30 km of a base station. Take a look at this map! Incredible.

SmartNet Aus Site Map | SmartnetAus

2-centimetre differential GPS would totally work in Victoria. Very very very cool. HW2 Teslas have an awesome source of redundancy for the multi-camera system. A car can place itself within the HD maps using cameras and its rough (1.5 m) GPS location, but with 2-centimetre GPS it can pinpoint its exact location as accurately as lidar. Amazingly cool.

Compare the base station map to the population map. Almost all of the populous areas are covered by base stations, and the more densely populated areas are more densely covered.

It still isn't enough to have high accuracy DGPS capability. First of all, the X centimetre figure is a 95% value. The other 5% must be accounted for, and if you look into it (on a per-site basis), you may find that it's not good enough for autonomous driving. The other problem is the integrity aspect. No one can know the accuracy of the DGPS navigation solution in real time, so an error bound must be placed on it. This bound would be derived based on the errors and threats affecting the system, e.g. receiver noise, GPS signal multipath, troposphere effects, ionosphere effects, equipment failure risk, transient environmental anomalies, etc... You'll find that when you try to factor all of these things in, your position error bound can go into the tens or hundreds of metres. Effectively what this means is that your DGPS solution may tell you that you are at a particular location with high accuracy, but for safety, your error bound might say that you are guaranteed to be in a position around there out to 50m. Differential GPS is not some magic bullet that you can use for tactical ground navigation, especially where safety of life is involved.

 

[strangecosmos]

In the first paper I linked to (the European one), every recorded error was below 4 cm out of 478 trials. A similar test in Australia found a lower level of accuracy, but still good. The average error for GPS using the RTK base station was 6.5 cm, with a standard deviation of 0.08 cm. Do you think there are common sources of error that they aren’t measuring? Or is this generally the level of reliability you could expect, barring equipment failure and other rare occurrences?

As a rule of thumb, I think of 10 cm as a significant threshold for autonomous driving because it’s approximately the limit of human driving accuracy, at least according to some parking data I found. So if you can measure with under 10 cm of accuracy, you theoretically have all the input you need to drive at least as accurately as a human.

Even 2 m of GPS accuracy is probably good enough for autonomous driving, I think. Combining HD maps with a multi-camera system and just a rough GPS location, the car should be able to pinpoint its exact location to under 10 cm using landmarks. It would be great to have a GPS accurate to under 10 cm as well, since it would provide redundancy to the cameras. But I don’t see that it’s necessary. It would just be cool to have that extra layer of redunancy.

I wasn’t seriously suggesting that autonomous cars should drive in zero visibility fog just relying on HD maps, high-precision DGPS, ultrasonics, and radar. I definitely think cars should stay put if they can’t see. I just think it’s a fun thought experiment to describe what’s theoretically possible. If high-precision DGPS can be accurate enough to guide a car clear from all fixed objects in the HD map and keep it in its lane, that shows the redundancy it can provide to cameras (which will always be the primary sensors).

It looks like there is a growing network of RTK stations in parts of North America. That’s very promising. It would be cool if we could get sub-10 cm GPS accuracy here. The more redundancy the better.

I also wouldn’t put it past Tesla to start setting up their own base stations at some point. Tesla loves vertical integration and being in control (e.g. Tesla Music). It would be in character for Tesla to start installing base stations on the roofs of its stores and service centres and at Supercharger stations. Maybe not in the near-term, but it would make sense to do it in cities when it launches the Tesla Network.

 

[AnxietyRanger]

the car should be able to pinpoint its exact location to under 10 cm using landmarks

Which is what I was saying in #5 and for some odd reason you disagreed with.

What GPS receiver do HW2/AP2 cars use? #5

The landmarks angle, really, is the most promising HD positioning angle at this time - and something where fleet-learning also might be useful.

 

[BinaryField]

In the first paper I linked to (the European one), every recorded error was below 4 cm out of 478 trials. A similar test in Australia found a lower level of accuracy, but still good. The average error for GPS using the RTK base station was 6.5 cm, with a standard deviation of 0.08 cm. Do you think there are common sources of error that they aren’t measuring? Or is this generally the level of reliability you could expect, barring equipment failure and other rare occurrences?

I looked at this Australia paper, and I will say that it is extremely important to quote results of academic studies in context. There are a number of details that stand out to me.

1. There was no truth source available for the test vehicle. The accuracy numbers that are stated are cross-track errors only, using the centre of the road as reference. This means that they are just error estimates for their proof of concept design. There is no way to get along-track errors. However, if the along-track errors are around the same size as the cross-track error you cited, then the vehicle could be 9cm away from where it thinks it is, on average.
   
2. The term "Real-Time Kinematic" is a bit abused in the navigation community. Here, they are doing a post-processed "RTK", so it isn't really real-time. The authors themselves talk about the complexity of the data processing, taking many hours to complete. High quality integer ambiguity fixing is a time-expensive process, and it isn't always successful.
   
3. The scope of this paper was to get accurate tracking of a vehicle in order to to study driving behaviour. This means that it is most important to get accurate positioning data, and it is not important to have integrity assurance or be done in real time. Autonomous driving needs all three.

As a rule of thumb, I think of 10 cm as a significant threshold for autonomous driving because it’s approximately the limit of human driving accuracy, at least according to some parking data I found. So if you can measure with under 10 cm of accuracy, you theoretically have all the input you need to drive at least as accurately as a human.

Even 2 m of GPS accuracy is probably good enough for autonomous driving, I think. Combining HD maps with a multi-camera system and just a rough GPS location, the car should be able to pinpoint its exact location to under 10 cm using landmarks. It would be great to have a GPS accurate to under 10 cm as well, since it would provide redundancy to the cameras. But I don’t see that it’s necessary. It would just be cool to have that extra layer of redunancy.

Integrity concerns aside, you'll have a very hard time blending a 2.0m-accurate source with a 10cm = 0.1m-accurate source. The former will have very little weight in the final solution. Differential GPS at that level is not suitable for tactical ground navigation.

It looks like there is a growing network of RTK stations in parts of North America. That’s very promising. It would be cool if we could get sub-10 cm GPS accuracy here. The more redundancy the better.

I also wouldn’t put it past Tesla to start setting up their own base stations at some point. Tesla loves vertical integration and being in control (e.g. Tesla Music). It would be in character for Tesla to start installing base stations on the roofs of its stores and service centres and at Supercharger stations. Maybe not in the near-term, but it would make sense to do it in cities when it launches the Tesla Network.

I don't think it will be happen for the reasons I've cited in my past posts. Maintaining high accuracy + high integrity navigation equipment is not cheap. The RTK network you cited is not to be used for safety of life applications. People would use those stations for surveying, for instance, where integrity is not a concern and the data will be cleaned-up in post processing.

A bit of a consolation is that it seems that the agriculture industry can easily take advantage of GPS technology for autonomous farming vehicles. In that context, it's probably more important to have accurate positioning over integrity, since presumably they would be operating in consistent environments (e.g. an open crop field). Quite a bit of work has been done in this area, so you may be interested in reading about it.

 

[strangecosmos]

Thanks for your insight on the paper.

As I understand it, the reason GPS is needed for HD maps is just this. The car checks for where it is geographically using GPS. It then uses its cameras to look for a nearby landmark that is mapped in its HD maps. It then uses visual SLAM, with the landmark as a reference point, to determine its location in the HD map to under 10 cm. At that point, you could theoretically just turn the GPS off. HD maps and visual SLAM will guide the car. Not GPS.

I guess if you think centimetre-scale GPS isn’t accurate or reliable enough then it can’t be used even for redundancy. Oh well. I guess that will save Tesla a lot of trouble and money, then.

 

[J1mbo]

A bit of a consolation is that it seems that the agriculture industry can easily take advantage of GPS technology for autonomous farming vehicles. In that context, it's probably more important to have accurate positioning over integrity, since presumably they would be operating in consistent environments (e.g. an open crop field). Quite a bit of work has been done in this area, so you may be interested in reading about it.

This sort of application is very mature and has been available on the market for years. Autoguide can be specified with "military spec" GPS for cm level accuracy.

 

[BinaryField]

Thanks for your insight on the paper.

As I understand it, the reason GPS is needed for HD maps is just this. The car checks for where it is geographically using GPS. It then uses its cameras to look for a nearby landmark that is mapped in its HD maps. It then uses visual SLAM, with the landmark as a reference point, to determine its location in the HD map to under 10 cm. At that point, you could theoretically just turn the GPS off. HD maps and visual SLAM will guide the car. Not GPS.

I guess if you think centimetre-scale GPS isn’t accurate or reliable enough then it can’t be used even for redundancy. Oh well. I guess that will save Tesla a lot of trouble and money, then.

No problem.

I see, then it might make sense if the car's GPS as it is now is used to initialise other more precise navigation functions. It just can't be meaningfully blended in with another navigation source that is a whole order of magnitude more precise.

Sort of the bottom line is that even if you have a system that is accurate to a centimetre scale, you have to take into account that accuracy figures represent nominal operation. Any safety of life system needs to protect the user against off-nominal conditions, e.g. a bad multipath event caused by GPS signals scattering off of the edge of a structure. When you start making those considerations, the performance of the system goes down.

This sort of application is very mature and has been available on the market for years. Autoguide can be specified with "military spec" GPS for cm level accuracy.

Thanks, really cool link. Are these vehicles actually autonomous, or do they require an operator to sit in and monitor the system? The difference is a pretty enormous leap in engineering.

 

[J1mbo]

Thanks, really cool link. Are these vehicles actually autonomous, or do they require an operator to sit in and monitor the system? The difference is a pretty enormous leap in engineering.

Semi. A driver is still needed to take over if something unpredictable happens - the system isn't designed to look out for pedestrians or broken ploughs