HW2.5 capabilities

[croman]

Yah, that's the untracked car problem, the car doesn't "see" the car sometimes if it's not moving and didn't see it beforehand moving. Never ever trust the car to stop unless, your behind a moving car and even then be very wary around stop / go traffic especially local traffic.

Not that anyone should trust AP2 but I've found it will consistently stop for untracked cars at 40mph or less. Anything above that and it will not stop (at least before I slam the brakes). Now, it will sometimes "see" the car and slow down only to lose it again and then accelerate and then SLAM on the brakes. Its not comfortable. Its quite unnerving but <40mph and its not bad (though sometimes unnerving still). That's untracked never before seen cars (waiting at a stop light).

If it tracked the car and disappeared from the IC, it is still tracking the car and will slow down consistently. But you can never trust AP.

 

[BeesKnees]

HW 2.0. 17.42. Sorry, left that off previous post.

 

[brkaus]

Not that anyone should trust AP2 but I've found it will consistently stop for untracked cars at 40mph or less. Anything above that and it will not stop (at least before I slam the brakes). Now, it will sometimes "see" the car and slow down only to lose it again and then accelerate and then SLAM on the brakes. Its not comfortable. Its quite unnerving but <40mph and its not bad (though sometimes unnerving still). That's untracked never before seen cars (waiting at a stop light).

If it tracked the car and disappeared from the IC, it is still tracking the car and will slow down consistently. But you can never trust AP.

I agree with you observations. I tend to “speed down” to about 30 with the TACC control as I approach those situations. And still keep foot ready to brake.

 

[BigD0g]

Not that anyone should trust AP2 but I've found it will consistently stop for untracked cars at 40mph or less. Anything above that and it will not stop (at least before I slam the brakes). Now, it will sometimes "see" the car and slow down only to lose it again and then accelerate and then SLAM on the brakes. Its not comfortable. Its quite unnerving but <40mph and its not bad (though sometimes unnerving still). That's untracked never before seen cars (waiting at a stop light).

If it tracked the car and disappeared from the IC, it is still tracking the car and will slow down consistently. But you can never trust AP.

I don’t play with it to much, it’s just not consistent and smooth to me, I just disengage and regenerate to the car. One day maybe....

 

[Padelford]

Question about AP2.0. Does anybody else find that in both TACC and AutoSteering, the car will maintain full speed as it approaches a stopped car ahead of it, until pretty hard braking is required? In this game of chicken, I nearly always hit the brakes before the Tesla does. This is bad for regen as well, as it requires significant brakes.

Yes, and I think this is a major problem for full autonomous operation. The vehicle should at least be using radar to detect there’s a stationary object ahead & start braking earlier than I’ve observed.

 

[ChrML]

Yes, and I think this is a major problem for full autonomous operation. The vehicle should at least be using radar to detect there’s a stationary object ahead & start braking earlier than I’ve observed.

The radar can't tell the difference between a stationary vehicle in the road, an overhead roadsign or a sodacan standing on the road.

But it can tell that it's a vehicle when it's moving because, well, it moves, roadsigns and sodacans don't.

That's why I think the cameras are the way of solving this, and not the radar. But they have yet to roll out that code.

 

[lunitiks]

The radar can't tell the difference between a stationary vehicle in the road, an overhead roadsign or a sodacan standing on the road.

But it can tell that it's a vehicle when it's moving because, well, it moves, roadsigns and sodacans don't.

This doesn't sound right. How do you know?
Bosch creates a map that uses radar signals for automated driving

 

[daktari]

Well if you think about it on a upward hill with a left or right turn, the car has no clue, since it can’t find the lane lines with the cameras facing the sky, they don’t pivot down like our eyes. This won’t get solved until they are leveraging the HD maps I suspect.

Not the case I am speaking about, similar to brkaus and other said though. Small undulations where the lanes are visible make the wheel turn, enough to be uncomfortable. Big crests, with a few seconds of invisible lanes will make the car steer in the ditch.

 

[ChrML]

This doesn't sound right. How do you know?
Bosch creates a map that uses radar signals for automated driving

It's how Doppler radars work. The receiver receives an overwhelming amount of data. By filtering out moving elements you can tell other vehicles on the road. The Tesla manual also states that the radar cannot see stationary objects.

Pulse-Doppler signal processing - Wikipedia

 

[chillaban]

It's how Doppler radars work. The receiver receives an overwhelming amount of data. By filtering out moving elements you can tell other vehicles on the road. The Tesla manual also states that the radar cannot see stationary objects.

Pulse-Doppler signal processing - Wikipedia

Automotive radar including the Bosch MRR is not pure Doppler. It’s both pulsed and frequency modulated. The radar object format provides recognition of individual stationary objects , except with such objects there might be a bunch of them in the scene, and it’s easier to code up a TACC system if you just ignore it altogether.

 

[Kant.Ing]

That's a myth and some excuse. The reason radars are mainly used to detect moving objects is because a stationary radar design cancels out the reflections from relatively stationary objects and such radar application considers only moving objects hostile. Given a good algorithm and computing power, radar can detect both kinds of objects and in a moving car, the radar subsystem cannot and should not differentiate moving and stationary objects.

 

[ChrML]

Automotive radar including the Bosch MRR is not pure Doppler. It’s both pulsed and frequency modulated. The radar object format provides recognition of individual stationary objects , except with such objects there might be a bunch of them in the scene, and it’s easier to code up a TACC system if you just ignore it altogether.

Yep! I believe Tesla mostly ignores the stationary objects alltogether. The radar itself will give info about them, but they are hard to interpret (causing false positives / phantom braking).

 

[chillaban]

Yep! I believe Tesla mostly ignores the stationary objects alltogether. The radar itself will give info about them, but they are hard to interpret (causing false positives / phantom braking).

They’ve been trying to push the radar sensor they use to its limits. In one of the AP2 threads, someone analyzed the radar binary blob that’s sent back as a part of most of the snapshot data. I believe it depicted 32 objects in the signature, and they were able to correlate it to trees, light poles, and all sorts of stationary data, fairly well localized.

Without a doubt the data is noisy and you can’t just make decisions off of that without the help of the fleet and a ton of software processing, otherwise you basically end up with the first 8 months of AP2 where it’d randomly slam the brakes at absolutely nothing.

With that said, completely ignoring stationary signatures is an underutilization of the sensor’s capability too. There’s arguably a lot of useful data in what’s the difference between what you’re seeing at a location right now vs what you saw in the past. That’s the kind of criteria that leds to “hey there’s a fairly large stationary object straight ahead that wasn’t there before...”

 

[lunitiks]

A quick follow-up on the "which Continental radar" discussion...

So we already know that Tesla pulls not one, but two CAN outputs from the Conti radar. Looking at the Model X wiring diagram, these channels are called "CAN" (+ and -) and "CAN2" (+ and -). In the Model S wiring diagram they're called "CAN Primary" and "CAN Secondary". Notice that Tesla does not use the word "redundant" for the second CAN-output, but "2" or "Secondary". In contrast, Tesla does use the word "redundant" on some other wiring connections, for example "Batt Redundant" for the redundant (duh!) Power Steering ECU power supply. (Of course these semantics don't rule out the possibility that the secondary CAN is in fact for redundancy, but it might indicate that it's for something else. Which I'll come back to...)

Anyway, both CAN channels are on the same physical connector (the radar only has one), and they both go directly to the glove box (AP2.5ECU).

Since we now have our first actual picture of the radar, it's possible to compare form factor with available Continental radar types. I have, and the one that stands out to me is the ARS 404-21. Among all Conti radars, this one is the closest match, I think. The "giveaway" is the small hole in the upper left corner of the radar, plus the indents in the "upper stage" of the radar. What do you think?

You'll notice that the connector seems to be different. And it is. The ARS404-21 comes standard with a Hirschmann connector, while Tesla uses Sumitomo. So Tesla's radar could be somewhat custom wrt. to connector and/or wiring. And that's where the two CAN channels come in:

According to Continental's datasheet and this very informative description, ARS404-21 apparently comes standard with only one CAN output. But, the Hirschmann 6-pin connector does have two pins for “CAN” that are “N.C.” – i.e. “Not Connected” (not in use). Tesla's Sumitomo-connector is also 6 pin, with GND, IGN, CAN+, CAN-, CAN2+, CAN2-.

Now the interesting part is that the ARS404-21 has software-adjustable range resolution and max range, and switches between Near and Far range:

Another benefit of the radar principle is a software adjustable range resolution and maximum distance. The sensor is capable to setting its range resolution from 1m and more up to 0.25 m, which can be done externally via CAN interface.
[…]
Selecting near range and far range scan is done via digital beam forming.

Furthermore, the radar:

can be set or changed by the customer for different functionality via CAN bus:
a. Sensitivity from low to high in several steps.
b. Objects (OB) or Cluster (CL)

Could it be that the two different CAN-outputs are for Near and Far range detection? (Custom for Tesla?)

 

[AnxietyRanger]

FWIW, the Germans have long used dual-mode radars (near and far) inside single units. Long being a decade or so.

Two outputs would fit with this.

 

[lunitiks]

Yes, and I think the «old» Bosch MRR Radar in AP1 and AP2.0 has some kind of dual mode too, with different FoV and range depending on mode.

I’m on real thin ice here, but could the difference, then, have to do with the «switching» between Near and Far? Could the two CAN outputs suggest that the Conti radar can give both Near and Far readings simultanously? Or might it perhaps have something to do with what I quoted above about Objects/Cluster switching?

Someone with a clue on radars please chime in. @DamianXVI you’ve analyzed the radar code, how do these things sound to you?

 

[Padelford]

Thanks, lunitiks, for providing the tech documentation on the Conti radar. Part of my background is in modern radar performance analysis, so this gives me lots to work with.

The radar in the Model 3 is an ARS408 based on the tech description document, which is good because the 408 has significantly better performance specs.

This is an impressive radar. It operates by emitting pulses with a frequency sweep (“chirp”) either up or down. The pulses can be relatively long for max Tx power, while the chirp provides enhanced down-range range resolution much better than the power pulse would provide.

This radar can do 1-D down-range imaging, and if the direction to the target varies in angle horizontally, the radar can do 2-D down-range imaging and target ID. This should help eliminating spurious returns, e.g., trees, buildings, etc. In fact, this is mentioned in the tech document on page 6.

The radar alternates short- and long-range scans. The two scan modes would have different Pulse Repetition Intervals (PRIs), and probably operate on different frequency channels (the radar has six available). Alternating the scans for the two range modes means the radar only needs one receiver chain which reduces cost.

I now want to pop out the frunk liner on my MX and try to find the radar!

 

[lunitiks]

@Padelford do you have a hardware 2.5 (brand new) car?

Also, how do you figure that the Model 3 has an ARS408, and not the 404? What am I missing in the tech documents?

 

[Padelford]

@Padelford do you have a hardware 2.5 (brand new) car?

Also, how do you figure that the Model 3 has an ARS408, and not the 404? What am I missing in the tech documents?

I have an AP 2.0 car.

RE the radar model, look at the package outlines in the Conti document, particularly the mounting flanges on the 404 unit. The Model 3 radar looks more like the 408.

 

[BigD0g]

I have an AP 2.0 car.

RE the radar model, look at the package outlines in the Conti document, particularly the mounting flanges on the 404 unit. The Model 3 radar looks more like the 408.

It appears you and @lunitiks need to compare some notes, I seriously doubt they are using two different radars for AP2.5 . my guess is they are the same, but I'm sure you all will figure it out.