HW2.5 capabilities

[AnxietyRanger]

I'm been having so much fun reading this thread, feeling so optimistic about the future of Tesla! Then suddenly a thought occurred to me, "where the heck are @AnxietyRanger and @Bladerskb?

I don't have an AP2.5 car, nor any technical access to these, so I don't have that much to contribute to a technical thread about it. And I'm happy there is no debate. I do read, this is one if my favorite threads too and on my notifications list. @verygreen is super-cool.

When we discussed AP2 vision, something where I could pitch in, I was all over it:

AP2.0 Cameras: Capabilities and Limitations?

e.g. AP2.0 Cameras: Capabilities and Limitations?
and AP2.0 Cameras: Capabilities and Limitations?
and AP2.0 Cameras: Capabilities and Limitations?

 

[AnxietyRanger]

Does sonar plus forward camera give you total coverage of the side for lane changes, or do you need the b-pillars too?

No, absolutely not. Ultrasonics in general are not fast enough and can't see far enough to be used for lane changes. Even if they'd solve the speed question (ultrasonics can be momentarily blind at certain speed deltas), they still can't solve the fast approaching car on the next lane without using the side repeater cameras (since they have no rear corner radars that most of the competition uses). Unless, of course, this being HW2.5 thread, some revision of HW2.5 (Model 3) actually has rear corner radars hidden there.

 

[mitchellh3]

So cameras 1 and 2 load main_narrow model - this much we know.
Now, cameras 3 and 8 load the fisheye model (so backup camera? really? I guess both are wide angle cams, but still....)
and now the kicker: cameras 4, 5, 6 and 7 are all loading the repeaters model (that used to be called "sides" in previous releases of the code) - that's really weird since the images you'd see from those cameras is kind of very different. I'd probably be less surprised if they had same model for main and repeaters I guess.

So this confirms they're loaded, right?

The next question I suppose is if they're running and sending any data back (snapshots or otherwise) to Tesla. Mainly: are they testing accuracy against more real world driving now in prep for some feature release later this year?

 

[calisnow]

No, absolutely not. Ultrasonics in general are not fast enough and can't see far enough to be used for lane changes. Even if they'd solve the speed question (ultrasonics can be momentarily blind at certain speed deltas), they still can't solve the fast approaching car on the next lane without using the side repeater cameras (since they have no rear corner radars that most of the competition uses). Unless, of course, this being HW2.5 thread, some revision of HW2.5 (Model 3) actually has rear corner radars hidden there.

Took you a bit to regroup but your job is safe. Whew. Wouldn't want you to miss a payment and have your car repossessed.

 

[AnxietyRanger]

Took you a bit to regroup but your job is safe. Whew. Wouldn't want you to miss a payment and have your car repossessed.

For the sake of this thread and its technical merits, anyone more who needs to insult me, may I at least request to do it e.g. here: Happy Birthday AP 2.0. Let's not derail this one.

 

[Reciprocity]

I'm been having so much fun reading this thread, feeling so optimistic about the future of Tesla! Then suddenly a thought occurred to me, "where the heck are @AnxietyRanger and @Bladerskb?

They are on the Audi forums talking about how awesome Audi is and how crappy Tesla is.

 

[AnxietyRanger]

They are on the Audi forums talking about how awesome Audi is and how crappy Tesla is.

Sigh. So much for my suggestion. Apparently personal insults are all the vogue on TMC these days.

 

[BigD0g]

Guys, let's stay on topic!!

 

[thomasmb]

So this confirms they're loaded, right?

Based on a quick look at the code he pasted, it doesn't verify it directly, yet. Depends on the ncameras variable value, which we don't know from that snippet. It just shows that the cameras are grouped to use the same NN.

I'm just happy to see that there has been som progress! I definitely was too optimistic when they announced AP2.0 and even in the time leading up to the summer, but it has been hard to be optimistic at all lately. Great to see something is happening at least, also noticing it while driving with .40 and .42.

As a side note; I'm just a bystander here, following it for the technical goodies like the ones @verygreen and @jimmy_d have presented us with (awesome work guys!). Trash talking other members is just ruining this thread IMHO.

 

[bjornb]

I am also a bit more optimistic now after we finally see some more substantial changes (not just driving algorithm changes which has been the case for some time until recently). Thanks to @verygreen and @jimmy_d for giving us insight/details!

Ultrasonic were mentioned above. Why are all ultrasonics that slow? I am aware that sound propogration is much slower than radar, but I have an impression the response time is in the hundreds of milliseconds region. There must be some kind of lag somewhere in the chain?

 

[BigD0g]

I am also a bit more optimistic now after we finally see some more substantial changes (not just driving algorithm changes which has been the case for some time until recently). Thanks to @verygreen and @jimmy_d for giving us insight/details!

Ultrasonic were mentioned above. Why are all ultrasonics that slow? I am aware that sound propogration is much slower than radar, but I have an impression the response time is in the hundreds of milliseconds region. There must be some kind of lag somewhere in the chain?

Perhaps to interpret the sound signals? My old remembering is the wind noise from different speeds affects ultrasonics and needs to be filtered out, kinda like noise canceling headphones, but this was very long ago, and they may have already solved for that.

 

[AnxietyRanger]

Isn't it very simple? Ultrasonics move at the speed of sound, whereas radar and vision move at the speed of light.

Slowness of ultrasonics is, anyway, the least of one's worries with automatic lane changes. Even if we assume the ultrasonics were fast enough in an AP2 Tesla (doubtful), their range is nowhere near sufficient to see a fast-approaching car on the other lane. That's what you optimally have a radar for (can see through trailing cars on your own lane) or alternatively you use a camera to do it.

Think of the fast-approaching ambulance, which ones do you see/hear first, its blinking lights or its sound...

 

[Saghost]

I am also a bit more optimistic now after we finally see some more substantial changes (not just driving algorithm changes which has been the case for some time until recently). Thanks to @verygreen and @jimmy_d for giving us insight/details!

Ultrasonic were mentioned above. Why are all ultrasonics that slow? I am aware that sound propogration is much slower than radar, but I have an impression the response time is in the hundreds of milliseconds region. There must be some kind of lag somewhere in the chain?

Yeah, at ~1000 feet per second, 15 feet each way is a trivial delay, something like 30 milliseconds total. Since they use a single crystal for sending and receiving, the response has to go back to the processing unit in real time.

I'm thinking that to reduce false positives and processing requirements the ultrasonic systems must be set not to react until they see several returns from about the same distance.

For all of which, the sensors on my X do pop up a bunch of false returns at times, especially on a windy day.

(The few automotive systems I found specs for in a quick search were promising detection in 300ms with 5 foot ranges, but none of them specified a pulse rate that I could see. In theory that could be up to 30 pulses, but I'm guessing it's more like 10 - and since Tesla is using longer ranges, the maximum possible pulse rate goes down.)

 

[brkaus]

Clarifying question? The title of this thread is HW2.5 capabilities... but if I understand correctly, the majority of the software dissection findings apply to HW2 & HW2.5? Just want to make sure I am on track.

 

[AnxietyRanger]

Yeah, at ~1000 feet per second, 15 feet each way is a trivial delay, something like 30 milliseconds total. Since they use a single crystal for sending and receiving, the response has to go back to the processing unit in real time.

I'm thinking that to reduce false positives and processing requirements the ultrasonic systems must be set not to react until they see several returns from about the same distance.

Fair enough, sounds like a plausible reason adding to it.

 

[bjornb]

Yeah, at ~1000 feet per second, 15 feet each way is a trivial delay, something like 30 milliseconds total. Since they use a single crystal for sending and receiving, the response has to go back to the processing unit in real time.

I'm thinking that to reduce false positives and processing requirements the ultrasonic systems must be set not to react until they see several returns from about the same distance.

For all of which, the sensors on my X do pop up a bunch of false returns at times, especially on a windy day.

(The few automotive systems I found specs for in a quick search were promising detection in 300ms with 5 foot ranges, but none of them specified a pulse rate that I could see. In theory that could be up to 30 pulses, but I'm guessing it's more like 10 - and since Tesla is using longer ranges, the maximum possible pulse rate goes down.)

If I remember correctly AP2 sensors have 8m range, which means up to 50ms response for a single pulse, and if you need to use several returns to filter out false positives and some processing time it will be slow..

I found a proposal to use ultrasonics for low speed AEB functionality here :
Ultrasonic Sensors in Urban Traffic Driving-Aid Systems

Anyway, as @AnxietyRanger mentioned above, the ultrasonics are useless for approaching cars because of the limited range alone.

Clarifying question? The title of this thread is HW2.5 capabilities... but if I understand correctly, the majority of the software dissection findings apply to HW2 & HW2.5? Just want to make sure I am on track.

Yes.

 

[BigD0g]

So riddle me this batman, if the ultrasonics are useless for fast approaching cars and we have no rear radars, nor good lighting in behind the car accept that which is provided by oncoming traffic.

How will the car judge distance of approaching cars in the night? I would think the side repeaters would be blinded by the oncoming traffics lights, much like our rear view mirrors if we didn't have the autodimming feature. So, how do the side repeaters determine distance of oncoming traffic at night?

I thought the ultrasonics were extended to 8m to help with this (2x the old version I think?)

 

[spottyq]

So riddle me this batman, if the ultrasonics are useless for fast approaching cars and we have no rear radars, nor good lighting in behind the car accept that which is provided by oncoming traffic.

How will the car judge distance of approaching cars in the night? I would think the side repeaters would be blinded by the oncoming traffics lights, much like our rear view mirrors if we didn't have the autodimming feature. So, how do the side repeaters determine distance of oncoming traffic at night?

I thought the ultrasonics were extended to 8m to help with this (2x the old version I think?)

Oh no, not again…
Please go here, there or even there. Or start your own thread.

And could we please leave this thread for the pro's ? They have already found and shared an awesome amount of information and I'd hate to see them discouraged by the constant derailing.

 

[jimmy_d]

No, absolutely not. Ultrasonics in general are not fast enough and can't see far enough to be used for lane changes. Even if they'd solve the speed question (ultrasonics can be momentarily blind at certain speed deltas), they still can't solve the fast approaching car on the next lane without using the side repeater cameras (since they have no rear corner radars that most of the competition uses). Unless, of course, this being HW2.5 thread, some revision of HW2.5 (Model 3) actually has rear corner radars hidden there.

Just realized that was a typo on my part - I meant to say repeaters, forward, and sonar. Was just wondering if b-pillar was critical to seeing slightly ahead and to the side, or if sonar can cover that adequately one lane over. Sounds like it's moot since @verygreen has discovered that b-pillars are getting used.

I suspected, but hadn't seen it written anywhere before, that sonar would be impacted by airflow. If you know of a good source on that topic could you share it?

EDIT: - looks like I jumped the gun on that request, I guess other people wanted it too.

 

[BigD0g]

Here's some data, not Tesla specific, but does give some good data on ultrasonics: https://senix.com/frequently-asked-questions/