Let's talk radar!

[lunitiks]

Warning: Nerdy-nerdnerdynerd thread

So what you're looking at is Tesla's forward radars; AP1, AP2.0 and AP2.5 respectively.
(OK, not 100% precise but close enough.)

Radar - Autopilot hardware 1 and 2.0 (Bosch)
Tesla Model S and X with Autopilot hardware 1 and 2.0 have a Bosch Mid-range radar sensor (MRR), Tesla part no. 1038224-00-B / 1057551-00-B (parts catalog link). Slight feature difference: Pre-Facelift AP1S has built in heating (leftmost picture), while Facelift AP1S, AP1X, AP2.0S and AP2.0X haven't because they're behind the front fascia (middle picture).

Bosch MRR product datasheet.

Feel free to google the part numbers - you'll get several eBay listings with nice pictures to look at

Anyway, the Bosch radar is connected to the Autopilot ECU (i.e. the MobilEye EyeQ3 chip on AP1, and the Nvidia-PX2-based APE board on hw 2.0), with a single pair of canbus wiring ("CAN1+", "CAN1-").

Radar - Autopilot hardware 2.5 (Continental)
Tesla Model 3, and S/X with Autopilot hardware 2.5 (built since ~August 2017), have a Continental ARS410 radar sensor. Tesla part no. 1108647-00-D.

So I can't find any official product datasheets on "ARS410" specifically, but if you look closely at the white print on the actual radar itself (rightmost picture), you find "IC: 4135A-ARS4B". Here's the datasheet for ARS 4-B. I don't know how much this datasheet can be trusted though, because Continental's website boasts somewhat better specs on ARS410.

Anyway, the Conti radar is connected to the Autopilot ECU with two pairs of canbus wiring ("CAN- Prim", "CAN+ Prim", "CAN- Sec" and "CAN+ Sec".)

Spec comparison
Now if I'm comparing the correct datasheets ( ), the Conti radar is:
- much more accurate at speed measuring (+/- 0.1kph vs Bosch's +/- 0.11mps)
- slightly longer detection range (170m vs Bosch's 160m)
- slightly lighter (130g vs Bosch's 190g)
- marginally less power consuption (4W vs Bosch's 4.5 w)

Questions unanswered:
- Why the dual canbus pairs on ARS410? Why not just two antennas on a single bus like MRR?
- Continental "max objects detected"?
- Cost difference?

Weight in, fellow internet slaves!

 

[OPRCE]

Great work on digging up the ARS-4B spec sheet! I'd been looking for that today with no luck ...

 

[iTech]

Warning: Nerdy-nerdnerdynerd thread

View attachment 384536

So what you're looking at is Tesla's forward radars; AP1, AP2.0 and AP2.5 respectively.
(OK, not 100% precise but close enough.)

Radar - Autopilot hardware 1 and 2.0 (Bosch)
Tesla Model S and X with Autopilot hardware 1 and 2.0 have a Bosch Mid-range radar sensor (MRR), Tesla part no. 1038224-00-B / 1057551-00-B (parts catalog link). Slight feature difference: Pre-Facelift AP1S has built in heating (leftmost picture), while Facelift AP1S, AP1X, AP2.0S and AP2.0X haven't because they're behind the front fascia (middle picture).

Bosch MRR product datasheet.

Feel free to google the part numbers - you'll get several eBay listings with nice pictures to look at

Anyway, the Bosch radar is connected to the Autopilot ECU (i.e. the MobilEye EyeQ3 chip on AP1, and the Nvidia-PX2-based APE board on hw 2.0), with a single pair of canbus wiring ("CAN1+", "CAN1-").

Radar - Autopilot hardware 2.5 (Continental)
Tesla Model 3, and S/X with Autopilot hardware 2.5 (built since ~August 2017), have a Continental ARS410 radar sensor. Tesla part no. 1108647-00-D.

So I can't find any official product datasheets on "ARS410" specifically, but if you look closely at the white print on the actual radar itself (rightmost picture), you find "IC: 4135A-ARS4B". Here's the datasheet for ARS 4-B. I don't know how much this datasheet can be trusted though, because Continental's website boasts somewhat better specs on ARS410.

Anyway, the Conti radar is connected to the Autopilot ECU with two pairs of canbus wiring ("CAN- Prim", "CAN+ Prim", "CAN- Sec" and "CAN+ Sec".)

Spec comparison
Now if I'm comparing the correct datasheets ( ), the Conti radar is:
- much more accurate at speed measuring (+/- 0.1kph vs Bosch's +/- 0.11mps)
- slightly longer detection range (170m vs Bosch's 160m)
- slightly lighter (130g vs Bosch's 190g)
- marginally less power consuption (4W vs Bosch's 4.5 w)

Questions unanswered:
- Why the dual canbus pairs on ARS410? Why not just two antennas on a single bus like MRR?
- Continental "max objects detected"?
- Cost difference?

Weight in, fellow internet slaves!

Maybe someone at Tesla though, "more is better right? sign me up". Now they're stuck with it.

Sounds crazy right?

 

[lunitiks]

Could the large difference in measuring accuracy (Conti ± 0.028 m/s, versus Bosch ± 0.11 m/s) be due to the different measuring schemes / operating principles? The MRR transmits frequency-modulated radar waves (FMCW; Frequency Modulated Continuous Wave), while the Continental ARS is a pulse radar, more specifically a pulse compression radar...

 

[ReflexFunds]

Does anyone have any idea what type of radar Tesla is developing in-house?

This project was reported in November but I never heard any additional details. Do you think it would be a heavily modified Continental ARS, or something Tesla is building from scratch?

Amir Efrati on Twitter

 

[lunitiks]

Don't know about that

However, sticking to what we do have (some) data on, I'm thinking since the AP2.0-radar inputs a single stream of canbus detection data to APE, while the AP2.5-radar inputs two streams - one "PRIMARY" (narrow, far range?) and one "SECONDARY" (wide, near range?) stream - then AP2.0 and AP2.5 must have different software processing of the radar detection data.

@verygreen? Anyone?

 

[HolyGrail]

Does anyone have any idea what type of radar Tesla is developing in-house?

This is pure speculation but someone mentioned Synthetic Aperture Radar (SAR).

I asked some radar guys and they doubt that such a sensor is close to series deployment or could even exist as working prototype.
A couple of years ago some professor built one as DIY project. Note that the radar still has to move perpendicular
along the scene to be captured. I could imagine that it might also work in moving direction, similar to 3d reconstruction with a monocular camera instead of a stereo camera.

 

[heltok]

One of my interest is SLAM, thus I subscribe to a few different youtube channels with related material. Last week an interesting video showed up in my feed about RADAR ego motion estimation:

I was surprised how well it worked. It seems to be good enough for being used in a backup to camera for safe emergency stops for getting an ASIL C to an ASIL D. This made me more confident that Tesla with some clever intelligence might be able to train their radar to safely stop if camera goes blind.

 

[strangecosmos]

Last week an interesting video showed up in my feed about RADAR ego motion estimation:

Holy cow! This is awesome. What about using radar to measure distance to nearby vehicles? What’s the best result you’ve seen on that?

 

[heltok]

Holy cow! This is awesome. What about using radar to measure distance to nearby vehicles? What’s the best result you’ve seen on that?

If there is a metal object in front of the car, you can be pretty certain that the radar will detect it. Not an expert on radars, but I assume distance accuracy will be decent like 2-10cm. The problem is usually angular accuracy. Latitudinal position error will depend on distance to object. Radar also natively gives information about the relative velocity of detections.

For reading up on radars:
http://www.iet.unipi.it/m.greco/esami_lab/Radar/automotive_radar.pdf

 

[Three3]

I’ve been following the radar situation somewhat closely—if Tesla is upgrading their radars I’d like to wait for that to happen before I purchase.

Continental teased their 5th generation radar at the end of 2017, claiming it would start production in 2019.

Fifth Radar Generation Meets Future Requirements for Automated Driving

I’ve been scouring the internet for the past few months looking for updates and coming up empty. Until that is I noticed this on their website today:

Continental Automotive

I don’t remember seeing this model on their site before, and even the picture is a rendering. The specs look quite a bit more impressive than their other radars.

@GigaGrunt and @Troy any thoughts on this? Seems a bit coincidental that this appears to finally have entered production on Continental’s side with the release of FSD/HW3...

 

[Troy]

@Three3, if they change the radar with HW3, lots of people who bought HW2.5 will be upset because Elon said there is no need to wait and the sensors are the same. Therefore Tesla has an incentive to leave other hardware changes to another time. Ideally, they should wait 6 to 12 months and then release HW3.5 with the new sensors if they have that kind of upgrade in the works.

 

[Daniel in SD]

I'd be curious to see a visual representation of what current automotive radar systems "see". Does anyone have a link?

 

[iTech]

I'd be curious to see a visual representation of what current automotive radar systems "see". Does anyone have a link?

Check out @verygreen's videos to get a sense of what the APU roughly sees.

 

[Three3]

Thanks @Troy - though to play devil’s advocate, and unless I’m mistaken, Elon Musk didn’t say that new vehicles wouldn’t include new sensors, just that FSD didn’t require them or entail any other retrofits to existing vehicles. So it could be that FSD works with the current sensors—and works slightly better with the new ones.

I do see the rationale however for them to wait on the new Continental radar, especially if they’re developing a new one in house.

 

[GigaGrunt]

Outside of my area of expertise.

 

[Three3]

Ah okay—thanks @GigaGrunt

 

[Soda Popinski]

Something I'm confused about. The older Bosch system boasts an elevation antenna, so the radar can distinguish between overhead signs and stopped vehicles (there's even a figure showing this). I've been lead to believe that AP1 and possibly the current versions can't distinguish between an already stopped vehicle and an overhead sign, leading to false negatives and the occasional running into firetrucks.

 

[lunitiks]

Something I'm confused about. The older Bosch system boasts an elevation antenna, so the radar can distinguish between overhead signs and stopped vehicles (there's even a figure showing this). I've been lead to believe that AP1 and possibly the current versions can't distinguish between an already stopped vehicle and an overhead sign, leading to false negatives and the occasional running into firetrucks.

Yes it can! The camera is pretty good detecting cars’ front and back. Firetrucks and sides of white semis, not so much

 

[astrowunder]

Do you know where the radar is physically installed? I've had a couple of times when I've needed to clean snow off the front to allow cruise control to work- it would be nice to know which part I need to clean instead of completely cleaning off the entire front of the car when it's blasting wet nasty snow!