Help me confirm the retrofit-ability of Autopilot 2.0 on the Model X

[Canuck]

That's an interesting thought. I myself knew very little two weeks ago. They can google around just like I did to pick up on things. I believe they know. They've likely been instructed by their management to act like they don't, to keep sales from being Osborned.

Given how many times in the past Tesla employees have provided information that is wrong, like saying no significant changes are coming before the "D", and no changes to the appearance before the refresh, and I could go on and on, I highly doubt they are being told of these changes but at the same time told to keep them quiet, or act like they don't know. You don't think that would get leaked? These are employees hired just like anyone else for similar types of jobs. Why would Tesla tell them in the first place? Also, most have no interest in going home and googling for answers that customers ask -- and where do they look if they did? Here? We have no answers. This is aside from the fact that the vast majority of Tesla employees don't own a Tesla and thus don't have the same interest as you and me. Rule #1 - don't rely on what a Tesla employee tells you.

Well at this point it seems moot anyway. I agree with many that with this significant update coming (one I totally did not see or think was possible), anything that looks like AP 2.0 would be years away.

Yes, I've heard the same thing about the P85 when so shortly thereafter the P85+ came out; then for the P85+ when so shortly thereafter the P85D came out; then for the P85D when so shortly thereafter the P90D came out; then for the P90D when so shortly thereafter the P90DL came out; then for the P90DL when so shortly thereafter the P100D came out.

I know AP is different but "years away"? This is Tesla we're talking about! The Model 3 is due out next year. I can't see it coming out with AP1.0 hardware so they need to get 2.0 on the S/X before it's release. Will it be some months away? Possibly. But years away? I doubt it.

 

[technomation]

Given how many times in the past Tesla employees have provided information that is wrong, like saying no significant changes are coming before the "D", and no changes to the appearance before the refresh, and I could go on and on, I highly doubt they are being told of these changes but at the same time told to keep them quiet, or act like they don't know. You don't think that would get leaked? These are employees hired just like anyone else for similar types of jobs. Why would Tesla tell them in the first place? Also, most have no interest in going home and googling for answers that customers ask -- and where do they look if they did? Here? We have no answers. This is aside from the fact that the vast majority of Tesla employees don't own a Tesla and thus don't have the same interest as you and me. Rule #1 - don't rely on what a Tesla employee tells you.



Yes, I've heard the same thing about the P85 when so shortly thereafter the P85+ came out; then for the P85+ when so shortly thereafter the P85D came out; then for the P85D when so shortly thereafter the P90D came out; then for the P90D when so shortly thereafter the P90DL came out; then for the P90DL when so shortly thereafter the P100D came out.

I know AP is different but "years away"? This is Tesla we're talking about! The Model 3 is due out next year. I can't see it coming out with AP1.0 hardware so they need to get 2.0 on the S/X before it's release. Will it be some months away? Possibly. But years away? I doubt it.

You should read up on "Pascal's Wager:"
A philosopher’s 350-year-old trick to get people to change their minds is now backed up by psychologists

 

[vandacca]

I've read the AP 2.0 release directly from TeslaMotors.com. There really isn't enough detail there to assess the practicality of this new approach to using radar, but I agree that it is a "non-trivial" implementation. It is a bit of a gamble to go down this route and hence there may be a chance of failure (small chance in IMHO).

I find this new direction intriguing and can't wait to see how it pans out. I really don't think they'll fail because there are so many ways to move forward with this new approach that they will surely find a successful path. They can always put more emphasis on video over radar if the radar feed is not as reliable as they expect it to be.

Also, the radar implementation is still supplemented by video and fleet-gathered conditions based on GPS locations, which will improve over time. This 3-prong approach is sure to succeed since Tesla can manipulate the algorithms to address for failures in any of the 3 systems.

However, I'm *very* skeptical about bouncing radar signals under cars to assess what's happening in front. I can't see that ever working, but the blog post was very light on implementation details. Not only that, I'm concerned about current vehicles not having enough processing power to perform all this analysis.

Having said all that, once they have this working for 1 radar, it would be significantly better to have 2 or 4 radars. The resolution would be much better and the system should be able to more quickly assess the situation around the vehicle. It would also need significantly more processing power, so I would expect AP2.0 to include additional processors.

Therefore, I would expect there to be new AP2.0 hardware in the future because it would significantly improve the system in 4 ways:
- better visual/radar coverage around the vehicle
- better resolution
- faster analysis (stereo radar should be able to assess dangers with fewer frames than a single radar)
- faster analysis via more processing power

But I don't see any benefit to AP2.0 hardware right now because so much development and data collection is still required. There is no point in adding more hardware until you get the existing hardware/software combination working. It would just complicate things. But it is possible that Tesla is further ahead than I'm imagining (doubt it). Once they learn from this first implementation, they should be in a better position to design/architect AP2.0 hardware. At least that's how I would approach it.

In summary, maybe AP2.0 hardware is further away than I had first expected. I wish I had more information on what's happening inside Tesla, but that's not how things work, so I'll just have to be patient.

 

[technomation]

But I don't see any benefit to AP2.0 hardware right now because so much development and data collection is still required. There is no point in adding more hardware until you get the existing hardware/software combination working. It would just complicate things. But it is possible that Tesla is further ahead than I'm imagining (doubt it). Once they learn from this first implementation, they should be in a better position to design/architect AP2.0 hardware. At least that's how I would approach it.

That's how I'd approach it too, but Tesla shouldn't wait too long to move forward with AP 2.0 hardware. NVIDIA provides a deep-learning development platform that's available today. The methods for collecting and analyzing more data exist and are being pursued by competing automakers.

 

[vandacca]

That's how I'd approach it too, but Tesla shouldn't wait too long to move forward with AP 2.0 hardware. NVIDIA provides a deep-learning development platform that's available today. The methods for collecting and analyzing more data exist and are being pursued by competing automakers.

I agree with you. However, I'm not too concerned with "competing automakers". This experimental radar point cloud implementation is probably only being pursued by Tesla (haven't heard of anyone else doing it). Also, existing automakers are going to be a lot more cautious about implementing this technology, since they tend to iterate at a much slower pace than Tesla.

Finally, is it a coincidence that there are new rumors out this week that Apple might be changing direction with Project Titan? The summary is that Apple is laying off workers and may provide a platform for other automakers (rather than come out with their own car).

But I too would like to see Tesla come out with AP2.0 sooner rather than later. Even if they don't use the new radar inputs right away, at least existing vehicles will have the hardware when the software is ready to use it.

 

[technomation]

This experimental radar point cloud implementation is probably only being pursued by Tesla (haven't heard of anyone else doing it).

What Tesla accomplished in its v8 software release with a coarse radar point cloud is indeed creative and unique, and it will buy them some more time to develop their follow-on AP 2.0 hardware platform. Elon has gone on record against LiDAR in the hardware suite because of its ineffectiveness in rain, fog, dust, and snow; he reiterated this objection even in his recent blog. All valid objections (in 2016), but I've never seen him also mention the prohibitive cost of LiDAR, $8K a puck (see Velodyne's website). At that price, Tesla couldn't have used LiDAR even if they wanted to.

That all may change with Quanergy's S3 LiDAR offering, which claims to have brought down the cost significantly with a small-footprint, solid state LiDAR device that Quanergy's founder further claims is not affected by weather conditions. Velodyne is also moving forward with their next-generation, much-lower-cost LiDAR. If Tesla's objections to LiDAR are truly being minimized, we can imagine that LiDAR might join Tesla's camera and radar hardware suite for AP 2.0 after all. LiDAR's ability to render a very fine 3D point cloud with more detailed zoom supporting object recognition, far beyond what radar can ever do, would add new capabilities to AP 2.0 that are not possible on AP 1.0. That's why every Tesla competitor---MB, Audi, GM, Ford, Lexus, Porsche, you name it---is pursuing LiDAR. Tesla may be able to stay ahead of them by augmenting its advanced radar with LiDAR, a better camera, and a more powerful deep-learning processor.

http://on-demand.gputechconf.com/gtc/2016/presentation/s6726-louay-eldada-quanergy-systems.pdf

 

[ohmman]

That's how I'd approach it too, but Tesla shouldn't wait too long to move forward with AP 2.0 hardware. NVIDIA provides a deep-learning development platform that's available today. The methods for collecting and analyzing more data exist and are being pursued by competing automakers.

Generally speaking, NVIDIA's deep learning back-end products are based off of their GPU history and are intended to parallelize and speed up the model build process. Because of the intense amount of processing power necessary to build a model, it's reasonable to assume that this particular hardware exists in a data center somewhere, and not on board the vehicle. The vehicle's processor is responsible for executing the outputs from the model, which is a far less intensive task.

An interesting thing about deep learning models is that, while intuition tells you that more features (generic term for input variables in machine learning) are going to be better, sometimes they're not. Deep learning in particular does a better job with an expanded feature set, but there can be limitations to that. So it's possible that adding more radar or visual data could deteriorate the current model. I doubt it, but it's worth remembering that more isn't always better.

 

[technomation]

Generally speaking, NVIDIA's deep learning back-end products are based off of their GPU history and are intended to parallelize and speed up the model build process. Because of the intense amount of processing power necessary to build a model, it's reasonable to assume that this particular hardware exists in a data center somewhere, and not on board the vehicle. The vehicle's processor is responsible for executing the outputs from the model, which is a far less intensive task.

Yes you're right, several years ago we were doing physics first-principles simulation work with NVIDIA GPU hardware in the data center of a global oil company. Now with the IoT wave, there's embedded intelligence, smart devices, edge or "fog computing" --in the car. From Quanergy's presentation below, seems that NVIDIA is going in this direction for autonomous vehicles (see slides 17-19). It's mindblowing how much compute power (a pair of Tegra X1's) they now have on these small-footprint cards going into AVs.
http://on-demand.gputechconf.com/gtc/2016/presentation/s6726-louay-eldada-quanergy-systems.pdf

 

[goneskiian]

I've read the AP 2.0 release directly from TeslaMotors.com.

Are you referring to the most recent "blog" post discussing the Version 8 software update?

 

[technomation]

An interesting thing about deep learning models is that, while intuition tells you that more features (generic term for input variables in machine learning) are going to be better, sometimes they're not. Deep learning in particular does a better job with an expanded feature set, but there can be limitations to that. So it's possible that adding more radar or visual data could deteriorate the current model. I doubt it, but it's worth remembering that more isn't always better.

Yes, I've seen some recent work out of U. Florida and elsewhere how deep learning on massive sets of disparate data can lead to a completely erroneous result. Still ML/DL has proven useful for visual recognition, which in my work, we usually cross-check with a multi-model approach. In the case of AVs, another model--perhaps the "radar model" in v8--could be used to validate DL results. Obviously significant development work is required to come up with a real-world reliable AV system, but I think they have the in-car platform to do the next-gen.

 

[Canuck]

You should read up on "Pascal's Wager:"
A philosopher’s 350-year-old trick to get people to change their minds is now backed up by psychologists

I'm very familiar with Pascal's Wager. What are principles for if not to be sold, right? While I don't believe in god, if you do, do you really believe that god is so gullible as to accept Pascal's Wager for entry into his/her kingdom? I would think any god deserving of worship would value those who ask questions, and would look down upon anyone so weak as to accept Pascal's wager, since it doesnt take much of an analysis to understand that it is the ultimate in religious hucksterism.

 

[u00mem9]

Seems like the triple-cam plan went the way of Mobileye.

 

[mikes_fsd]

Seems like the triple-cam plan went the way of Mobileye.

It could have... given that it is mobileEye's area of expertise!

 

[vandacca]

Are you referring to the most recent "blog" post discussing the Version 8 software update?

Yes, that's the blog post I was referring to.

 

[krazineurons]

@FredLambert i have been meaning to ask you this question ever since the v8.0 announcement went off. In your article Tesla Autopilot 2.0: next gen Autopilot powered by more radar, new triple camera, some equipment already in production you have written

Sources with knowledge of the Autopilot program told us that the new suite will keep the current front-facing radar and add more around the car, likely one in each corner. Additionally, the system will feature a new front-facing triple camera system for which we are told Tesla started installing new housing in the Model S production this week.

Does any of that change with the recent announcement?

 

[aesculus]

Generally speaking, NVIDIA's deep learning back-end products are based off of their GPU history and are intended to parallelize and speed up the model build process. Because of the intense amount of processing power necessary to build a model, it's reasonable to assume that this particular hardware exists in a data center somewhere, and not on board the vehicle.

This sort of stuff is starting to drift into non model areas too. Anything that needs math and parallelization seems to be ripe.

Kinetica (software) - Wikipedia, the free encyclopedia

 

[fengshui]

I do wonder where Tesla would put additional radars on the existing Model X. I count visible 12 radars already (6 each front and rear), plus the hidden radars in the door for the falcon wings. I could see them wanting a different model of radar, or radars optimized for different ranges, but they have nearly 360-degree coverage today.

 

[vandacca]

I do wonder where Tesla would put additional radars on the existing Model X. I count visible 12 radars already (6 each front and rear), plus the hidden radars in the door for the falcon wings. I could see them wanting a different model of radar, or radars optimized for different ranges, but they have nearly 360-degree coverage today.

The "12" radars you count are not radar, they are the ultra-sonic sensors. They are short-range (e.g. Less than 20').

There is only 1 radar, and it's in the camera housing that the rear-view mirror is attached to. This has a much longer range.

Rumor has it, at one point Tesla was considering having 4-5 radars, the existing one plus one on each corner of the car.

 

[Saghost]

The "12" radars you count are not radar, they are the ultra-sonic sensors. They are short-range (e.g. Less than 20').

There is only 1 radar, and it's in the camera housing that the rear-view mirror is attached to. This has a much longer range.

Rumor has it, at one point Tesla was considering having 4-5 radars, the existing one plus one on each corner of the car.

All correct except for one detail. The single front facing radar isn't in the camera housing, it's in the front valence somewhere. My personal theory is that it is right behind the little depression in the nose of the X below the T symbol, but it could be elsewhere under the front - not necessarily on the centerline, either.

 

[fengshui]

All correct except for one detail. The single front facing radar isn't in the camera housing, it's in the front valence somewhere. My personal theory is that it is right behind the little depression in the nose of the X below the T symbol, but it could be elsewhere under the front - not necessarily on the centerline, either.

Ahhh, thanks!