Autopilot in DMV

[Damien Sandow]

I will be picking up my Model S on Friday if all goes to plan and I couldn't be any more excited. I live in Northern Virginia (Alexandria), but work several different places in the area. Twice a week I am over in Montgomery County MD, twice a week in Kingstowne, Once a month in Odenton, and once a week from home. Therefore I will be taking quite a few different roads and fully intend on taking advantage of Autopilot.

With that being said I wanted to create a thread and talk about folk's experience with Autopilot in the area. Basically this could be a place where we can put watch out areas where the software struggles or has consistent problems, in order to keep one another safe. Are there any spots that I should avoid or be extra cautious?

Thanks!

 

[MGMDaware]

First spot that comes to mind is on ramp that connects 395 s to 295 s right in front of the MGM. My AP2 always loses it there. The lane markings are a mess. Also, 395 n in downtown DC.... The fast lane has a metal grate parallel to the lane markings that confuses the cameras and tends to make the car hug the right most lane. Makes me nervous every time.

 

[Saghost]

Watch for places with bad lane markings. Watch for places objects are near the lane edges, and places with oddly shaped objects overhead.

 

[boaterva]

One place on my ‘must test’ list that still fails (and I’m sure there are many like it) are where a multi lane road comes into another road perpendicularly. And you have three lanes of curved ‘guide lanes’ of turning markers to help the other road’s traffic stay in their lanes.

Autosteer seems to get confused sometimes when seeing these when going straight ahead on the first road and thinking it’s supposed to follow the curves over to the next lane over. It doesn’t see ahead and just go straight.

My spot is on 50 West at Stonecroft in Fairfax County (almost to Loudoun) but I’m sure there are other cases of this. I’ve reported it but not fixed yet.

I think the real issue is that the intersection is so wide (over six lanes total with buffer, etc.) that the camera can’t see far enough ahead to the straight continuing lanes of 50 downstream and fixates on the curves coming in from the ‘guide lines’ of Stonecroft at the intersection.

 

[Saghost]

Oh, and if you've been reading a bunch of threads this might be obvious, but most importantly:

Watch for stopped vehicles the car never saw moving, especially if they are of unusual shapes or at odd angles or partly in the lane!

Tesla has one of the best adaptive cruise systems in the business, but it can still struggle with recognizing stationary cars, something that's hard for all the systems to do and most do a lot worse than AP/TACC does.

 

[Damien Sandow]

Watch for places with bad lane markings. Watch for places objects are near the lane edges, and places with oddly shaped objects overhead.

As I've been driving around lately I've been looking at things like this. It really seems like there are a lot of places with terrible lane markings, especially on 395. It is probably best to be mostly hanging out in the center lanes with all the on and off ramps in the area I imagine.

 

[Saghost]

As I've been driving around lately I've been looking at things like this. It really seems like there are a lot of places with terrible lane markings, especially on 395. It is probably best to be mostly hanging out in the center lanes with all the on and off ramps in the area I imagine.

One of the things AP1 (and presumably AP2+) has gotten much better at over the last year through the various firmware updates is handling the holistic path situations - these days it can usually find the right place to be for thirty seconds or more even though it isn't seeing either lane line or a car to follow - it used to veer away in the first five seconds or so.

So yes, it's certainly something to watch (and hold the wheel for,) but I've been pleasantly surprised at how well it does a lot of times lately.

 

[Mirlen]

I wish the car would hug the left lane marker on unmarked onramps where the 'lane' gets 'wide' from the merge. Right now it tends to center itself in the 'wide' area and wobble and I feel a bit odd driving like that I've also had a few incidents of emergency braking kick in when going under large 'across-all-lanes' signage on 202 around King of Prussia and in one overpass area. This was a problem a long time ago when I first got the car and it seems to have regressed in the last couple versions and come back...makes me a little nervous using AP in those situations.

Firmware: 2018.14.2 a88808e installed 4/17/18

 

[ArbitrageMan]

Does anyone take the exit from 495N to 123E in Tysons? That ramp is backed up badly every day. How does AP handle it?

(My car is on order. Can't wait to let it deal with that ramp.)

 

[sdugan99]

Most of my AP driving is Howard county to Hyattsville. Three main routes 95 to 495, 295 to 495, and 100 to 2 to 50(daughters school is in severna park) I am on AP for the Majority of this commute each day except with minor exceptions, really only when AP doens't detect the exit lane for me to use the blinker to get over. But I have noticed it is getting much better for slowing speed into corners, still comes out a little to fast. There were some tricky spots but it seems to have improved. Only thing I still get about once a week is the phantom slam on the brakes for now reason. I have taken the right hand land of rt 2, to the off ramp merge with traffic on 50 with autopilot engaged. I have to go to Tysons next week I will keep on eye out for tricky spots.

 

[Damien Sandow]

Much appreciated everyone! Tomorrow is the big day, and I'll be sure to share anything useful after some driving in the area.

 

[kevinf311]

295 in... well, everywhere, tends to have terrible lane markings.

That being said, so long as the traffic is stop and go (which is often) my AP1 car seems to handle things very well. The car is relying more on the car in front of it most of the time (based on what I see turn blue on the IC) and the crappy lane markings don't really come into play. When traffic is moving fast enough to require using the lane markings, the car does a decent job of figuring out what is supposed to be a lane (sometimes heuristically as mentioned above, sometimes due to the edge of 295 being high contrast) but depending on the car in front of you the transition between the two can be a little wobbly.

That's really my only complaint with AP (for AP1 at least); your car drives more or less the way the car in front of you is driving. If they're smooth and trying to absorb traffic accordions then so are you. If they are real hard on the accel/brake hugging the car in front of them as they move 2 feet at a time... then so are you. I don't have an elegant way for this to be improved, so I don't ding Tesla too much on this behavior. A slower acquisition rate during stop go might help but I also remember back after AP was launched that everyone was complaining about how slowly the cars regained the target speed in traffic (causing folks to try and pass them only to have the car eventually be going fast again). Bracketing the acquisition rate into low speed and high speed might be possible but not having real insight into the inner workings of either AP1 or AP2 I don't know how feasible or effective it would be.

Another road feature to keep an eye out for that highlights the limitation of the forward camera(s) is the gentle hill with a slight curve. You can watch the road marking display fade towards your car avatar on the IC as the camera (and you) become unable to see markings over the crest. Where AP and the driver differ, however, is AP has no reason to believe the road doesn't continue straight whereas the driver is probably a little more clued in. This typically results in an angrier style of lane wobble as the camera angle comes back down to see the continuation of road markings. Keeping an eye on the road markings "drawn" on the IC is an excellent way to assess what the car is about to do on the road. There is SO MUCH data being fed to the driver in AP cars (even without AP enabled). Not to get on a soap box but if a driver is monitoring the AP indicators on the IC in addition to the actual road conditions, and is cognizant of the delta between, then there is ZERO reason that an incident should occur, ever (regardless of the live-wire issue of wheel holding). Any accidents past that level of awareness are truly that, accidents.

Back on topic, the car is randomly amazing in inclement weather. With the caveat that inclement weather is an inappropriate time to rely on AP, in observing what the car thinks it sees via the IC (and with some testing where safe), sometimes the car is seeing lane markings when I am unable to due to glare or other environmental impediments. While impressive the consistency in lane acquisition is not high enough for me to feel comfortable leaving the car in AP during these times.

I'd be interested to hear from any folks that have driving with AP in an area that employs the use of "Bott's Dots" (I don't know of any in the DMV however) and whether the car has an easier time defining lanes with those over the small single reflectors used here on the Beast Coast.

 

[Saghost]

That's really my only complaint with AP (for AP1 at least); your car drives more or less the way the car in front of you is driving. If they're smooth and trying to absorb traffic accordions then so are you. If they are real hard on the accel/brake hugging the car in front of them as they move 2 feet at a time... then so are you. I don't have an elegant way for this to be improved, so I don't ding Tesla too much on this behavior. A slower acquisition rate during stop go might help but I also remember back after AP was launched that everyone was complaining about how slowly the cars regained the target speed in traffic (causing folks to try and pass them only to have the car eventually be going fast again). Bracketing the acquisition rate into low speed and high speed might be possible but not having real insight into the inner workings of either AP1 or AP2 I don't know how feasible or effective it would be.

Follow distance settings make a big difference here. 1 is marginally inside the car's apparent response loop, producing a harder braking reaction than the car it follows. 2 is borderline, matching the car in front pretty closely. 3 and up are much better, with smoother braking events as the car uses the extra following distance dynamically.

Of course, 4 and above tend to result in folks cutting in or even going around you and back in - they're all so eager to follow someone to closely and be part of the next accident (Honestly, I'm only a little better when I'm in control. It's a bad habit most of us seem to have gotten in to that you've gotta watch out for.) Of course, some idiots will cut in at any setting. At anything below 4, there's a good chance you'll need to break out of AP to avoid a collision when they do it - the car doesn't react to them until they're more than halfway in the lane in most cases.

In theory, the car should have a much tighter reaction time than it acts like - the radar is feeding distance data 20 times per second and velocity data potentially much more often than that (continuous wave radar always has doppler shifted signal coming back,) and the iBooster can hit wheel lockup in 125 ms (1/8 s) and presumably more normal braking events much quicker.

Instead, it feels like it has a half second to one second actual reaction time in practice - between seeing things happening in front of the car in front of me (which the radar is supposed to be doing now too,) and the difference in reaction times, AP reacts a lot slower than I do as a driver, which is part of why it's so rough unless you give it a big following distance. I mostly run it on 3 or 4.

 

[Damien Sandow]

In my limited time with the car I am getting cut in front of using 1 as the follow distance

 

[kaczmarczyk]

In my limited time with the car I am getting cut in front of using 1 as the follow distance

I use 5 on the highways and route 28. That seems good enough for my car to react to cars merging /cutting in front of me.

As others have said, be aware when they are iffy lane markings, on /off ramps, stationary objects. I have the most confidence when I am following another car. Autopilot is a God send around the metro area given our traffic. Pat attention and enjoy your Tesla!

 

[kevinf311]

Follow distance settings make a big difference here. 1 is marginally inside the car's apparent response loop, producing a harder braking reaction than the car it follows. 2 is borderline, matching the car in front pretty closely. 3 and up are much better, with smoother braking events as the car uses the extra following distance dynamically.

Of course, 4 and above tend to result in folks cutting in or even going around you and back in - they're all so eager to follow someone to closely and be part of the next accident (Honestly, I'm only a little better when I'm in control. It's a bad habit most of us seem to have gotten in to that you've gotta watch out for.) Of course, some idiots will cut in at any setting. At anything below 4, there's a good chance you'll need to break out of AP to avoid a collision when they do it - the car doesn't react to them until they're more than halfway in the lane in most cases.

In theory, the car should have a much tighter reaction time than it acts like - the radar is feeding distance data 20 times per second and velocity data potentially much more often than that (continuous wave radar always has doppler shifted signal coming back,) and the iBooster can hit wheel lockup in 125 ms (1/8 s) and presumably more normal braking events much quicker.

Instead, it feels like it has a half second to one second actual reaction time in practice - between seeing things happening in front of the car in front of me (which the radar is supposed to be doing now too,) and the difference in reaction times, AP reacts a lot slower than I do as a driver, which is part of why it's so rough unless you give it a big following distance. I mostly run it on 3 or 4.

I end up switching my follow distance around a lot due to the nuances you describe. Follow distance set to 1 is more or less a "polite MD follow distance" it seems, anything greater is an invitation to other drivers to cut in (sans turn signal of course )