Autopilot 7.0: When to be most on your toes?

[Papafox]

We can make the transition to AutoSteering safer for the Tesla community if we share information about the current potentially troublesome areas of the autopilot system so that we can learn from others and not have to learn every lesson by trial and error. I also want to post some Driver Beware cautions that may not be related to any weakness in the autopilot but are issues that could potentially affect safety of driving with autopilot engaged. Software update 7.0 gives us great tools to use in our driving. Let's be smart in how we employ them. I'm hoping you and others will contribute to this knowledge base. I hope to consolidate some subsequent contributions in this first post so that the information isn't too spread out.

Most important: Tesla says autosteer is only to be used when a center divider is present. Please find language in this post. You cannot expect autosteer to keep you safe without a center divider if Tesla prohibits such an operation.

Let me continue with a known rough-spot: off ramps.
Off ramps from straight highway segments- On a 3 hour drive on Interstate 80 from Reno to Sacramento and return today I passed many such off ramps while in the right lane. Typically, my Model S veered slightly toward the exit and then corrected back ok. People traveling on other highways with different markings may have had different experiences. If so, speak up.
Off ramps at the beginning of a left turn of the highway
This type of off ramp was problematic for the autosteering. The car is going straight, the off ramp continues to go straight, but AutoSteer needs to begin a left turn at the proper moment. See photo below.

Off ramps at the end of a right turn of the highway

Off ramps at the end of a right turn of the highway can be problematic. The car is turning right, the off ramp is positioned well to accept the car on its current trajectory and now the highway veers left of the current path.

Lanes that split

Lanes that merge

Lanes that are marked by a series of white raised circles (botts_dots), rather than lines (reported by a friend in Hawaii and confirmed by mainland users as well)

Lines painted on highway that compete with lane marker lines for directing AutoSteering (see this post in this thread)

Lack of edge lines on highway (solid yellow on left side of left lane, solid white on right side of right lane) Look for these.

Southern California HOV Lane:
A thread participant reports here: "... Its a double yellow which changes to dashed white when you are allowed to enter and exit the lane. My car wanted to exit most of the times the white dashes were transitioning back to the double yellow."

Wet Pavement at Night with Glare: Sometimes this situation interferes with autopilot according to this post.

Driver Beware:

Driver Beware: People on the highway
I came over a hill today on Interstate 80 to discover a vehicle parked on a narrow shoulder, its door opening into my lane and someone jumped out. Fortunately, the other lane was open and I hand-steered into the other lane. You would waste far too much time asking autopilot to bring you into the other lane. Turn the wheel, the AutoSteering will kick off.

Driver Beware: Challenge piled on top of challenge
If you pile one challenge on top of another challenge on top of another, you make a more difficult driving solution for man or autopilot. I suggest you getting comfortable with your autopilot performance and learn how responsive it is before really giving it difficult solutions that includes multiple simultaneous challenges such as:
* sharp or high-G turns at fast speeds
* narrow lanes
* lanes where barriers give little room on one side of the lane
* lanes where a semi-truck or other vehicle is crowding that lane from the adjacent lane or even crossing over the lane markers or swaying from side to side
* poor lane markings
* wet surfaces

Example of challenges piled atop challenges (submitted by Cyberax): narrow lanes, bus in adjacent lane and crowding the Tesla's lane, lane marked by white dots instead of lines.

Need further Input:
* Ability of sensors to pick up buses and large trucks reliably when abeam the Tesla. Some drivers believe that shadows created by large trucks might make lane markings between truck and Tesla difficult for Tesla to see. Please look to see if blue lane marker on autopilot dashboard indicator disappears when next to truck. In meantime, please do exercise caution when abeam a big truck.

 

[aja2460]

Autopilot 7.0: When to be most on your toes?

For the next year or two...

 

[Cyberax]

* Driving beside a bus or a trailer truck. AP sometimes has problems with detecting it.
* When a slower vehicle merges into your lane. AP does NOT track cars in other lanes efficiently. It's almost as bad as TACC before it.

 

[Papafox]

* Driving beside a bus or a trailer truck. AP sometimes has problems with detecting it.
* When a slower vehicle merges into your lane. AP does NOT track cars in other lanes efficiently. It's almost as bad as TACC before it.

Cyberax, when you say that autopilot doesn't detect trucks and buses sometimes, do you base this upon the visual spherically shaped images that are supposed to appear on the side of the car that is displayed just below the speed, or are you basing this statement on something else? It'd be interesting to get other people to verify your observation.

 

[Cyberax]

I was driving on SR-520 (Seattle) bridge today - and AP almost collided my cur into a bus at the adjacent lane. Lanes there are narrow, (and delineated using the raised dots, not lines) and the bus was at the very edge of its lane. I was slowly overtaking it and I was watching both the IC and the road - there was no indication on the display that the AP sensed that the lane to my right was occupied. And then it started to veer right (as it generally likes to do).

The car actually beeped the side collision warning _after_ I took over.

I've seen the same happen with trucks as well - they are not reliably detected. I think the high ground clearance of buses and trucks foils the side sensors into thinking that there's more space to the right.

 

[nanimac]

Thanks Papafox for adding the HAWAII dot issue..others on the mainland have reported the same issue too!

 

[RDoc]

* Driving beside a bus or a trailer truck. AP sometimes has problems with detecting it.
* When a slower vehicle merges into your lane. AP does NOT track cars in other lanes efficiently. It's almost as bad as TACC before it.

I've also seen the problem with cars merging from the right. I wonder if the autosteering system doesn't have a way to predict the other vehicles path, that is, it only looks at the instantaneous relative position rather than it's trending position?
This seems to me to be a serious problem that should get fixed.

 

[stevezzzz]

Kudos to Papafox for starting this thread; I was feeling the need for something like it and went searching, so I'm happy it already exists.

Highway exits and lane splits deserve to be high on the list: nearly every reviewer, myself included, has noted the problems:

Firmware 7.0 - Page 62

There's a YouTube video going viral right now (titled something like 'Tesla Autopilot tried to kill me') that shows how fast things can go pear-shaped if a driver ignores the system's struggles in difficult conditions. To me, the lesson there is that the way Autosteer manages the handoff back to the driver's manual control is a very important piece of the puzzle, and needs improvement.

Autopilot tries to turn car against incoming car

 

[garygid]

This 7.0 version (2.7.56) version of AutoPilot appears to be a simple merger
of the old TACC and a new Lane Steering function. On many well-marked
freeways, where the curves are gentle, the speeds are appropriate for the
road situations, and the lane markings are easily-recognized lines on both
sides of the lane, the steering seems to do a rather nice job: smooth, gentle,
and keeping the car well centered.

All other conditions should be considered suspect, with hands instantly
ready to take over the steering.

Situations, like exeting a freeway onto a curvy, low speed exit ramp are
not handled by any automatic speed control (reduction), so it is usually
not appropriate to leave the TACC in control in this situation that normally
requires much reduced speeds to safely negotiate steering around the curves.

So, situations that required manual speed control, still require manual control.

Even though the steering function might evolve to calculate the
turning radius rather than just react to being off-center (which makes
the steering control a bit late on sharper curves), it would need to
learn to reduce speed when necessary to handle the aproaching
road conditions. But, for now, the TACC seems to be blindly determined
to go as fast as it can whenever it does not detect an in-lane vehicle ahead.

So, use the new features VERY cautiously, realizing that the TACC (and steering) were
only designed for a very limited subset of driving conditions, where no/few surprises
are expected. Please consider SAFETY first.

 

[stevezzzz]

There is at least one common situation where TACC does limit speed below the set point: while accelerating to the set point and turning at the same time, TACC reliably stops accelerating until the turn rate is reduced by the driver.

 

[Papafox]

Cyberax: I've used your situation as an example of multiple challenges in the first post of this thread. I've also asked for more input on sensing buses and large trucks abeam the Tesla

Stevezzzz: Thanks for the heads up about this incident. Looks like the driver was entirely too relaxed driving on autopilot with no divider between oncoming traffic in a very difficult autosteering situation.

Garygid: Well put!

Let's start looking at other clues as we discuss autosteering functioning. Stevezzz brought up the lack of blue lane markings in another thread, so let's whenever possible refer to the indications that the Tesla display gives us. From what I can see, the display thinks out loud by drawing semi-circles when it detects nearby objects, with grey being the first indication, yellow being the caution distance, and red being the warning distance. Whenever possible, please state display indications at time of autosteering challenge. See image below.

 

[mspohr]

There is at least one common situation where TACC does limit speed below the set point: while accelerating to the set point and turning at the same time, TACC reliably stops accelerating until the turn rate is reduced by the driver.

I've also noticed the AP limiting speed when driving around a sharp bend. I took a drive over Mt. Rose (a steep, winding road over a 9000 ft. pass) where the speed limit is 50. I had the AP speed set to 60 and noticed that on some of the hairpin turns, the AP lowered the speed to as low as 35 (which I felt was too low for some of the turns) and on others it lowered the speed but not as much as I would have liked.
My conclusion is that the AP does adjust the speed for curves. It may not do a very good job of estimating the "right" speed but hopefully this will improve as more data gets into the system.

 

[cmu5p]

TRACTOR TRAILERS: I routinely drive on a stretch of interstate heavily populated by tractor-trailers and have noticed auto-steer turning off a couple of times when next to or in the shadow of a tractor trailer. It has always seemed to be on my right. I think the shadow maybe obscures the lines? Will keep looking for patterns. It surprised me there was no warning, just 'ding-ding'-> autopilot off. Am pretty vigilant so I am hands off, but at the ready to take control anytime. Still pretty amazing technology.

 

[Papafox]

Here's another situation to beware of. On I-80 Westbound from Reno to Auburn, Ca., Exit 139 provides an unusually tough solution for autosteer. Notice that two sets of white lines exist: the regular road lines and a thinnner set of white lines that leads the driver towards the turnoff, which is a hairpin turn and cannot be taken at highway speed. My Tesla detected that something was wrong, handed control back over to me, and I made a small correction to the left, back into the highway lane and away from the turnoff. This is precisely the type of setup that needs either attention by Caltrans to get rid of that second set of white lines or some type of exception transmitted to the Tesla through a Tesla database, to provide the right solution to the problem of multiple white lines on the highway.

Notice one of many thinner white lines in a row, diverging to the right from the highway lane marker

In this photo, you can see additional white lines trying to lead the Tesla right towards the hairpin turnoff. Fortunately, the Tesla AutoSteer had already turned control back over to the driver at this point and the driver was ready to act if AutoSteer didn't disconnect

 

[dsm363]

It doesn't like roundabouts either.

 

[Mjølner]

Best ever dsm363 :tongue:

 

[johnbro23]

Good thread.

Doesn't tesla already have all this data on how people typically drive? It has a fleet of 200k cars doing a few million miles / day. It doesn't seem like Tesla has factored in enough of this GPS-related date to AP, but seems like it should be. Can someone with more of a technical background explain what is the holdup?

 

[brec]

... (see page 2 of this thread) ...

Thread page numbers are unreliable because of forum Settings::General Settings::Thread Display Options::Number of Posts to Show Per Page

To refer to a specific post, a link to it is best.

 

[Papafox]

Thread page numbers are unreliable because of forum Settings::General Settings::Thread Display Options::Number of Posts to Show Per Page

To refer to a specific post, a link to it is best.

Done. Thanks.

 

[Khatsalano]

Good thread.

I do think lane-biasing has a lot of potential. For example:

- detect vehicles that are laterally near you ... lane bias away from them.
- detect large semis or trailers that have high sway ... lane bias away from them.
- in a curve, lane bias should be towards inside of the curve as you enter, then fan out to center. Except when:
..........in a curve to the right and you're in the leftmost lane with room, lane bias to the left throughout the curve, especially with a car to the right
- when in the rightmost lane, deference should be given to the left lane marking, not the right one due to exits
- in overtake lanes, the car should know how to at least follow the slower vehicle in front, or, choose the rightmost lane. Right now, it just freaks.
- when in the leftmost lane and there is sonar detectable k-rail, the car should still bias left instead of right so it stays away from traffic. Barriers don't jump out at you, but other drivers do. Even centered is better than the current right bias. Exception when the barrier is closer than say, 4 feet.