Autopilot TACC - still a work in progress

[breser]

We were in a highway stop and go situation today and I had to override the TACC. Traffic had speeded up to about 60, then quickly, but not unusually so, slowed down and stopped. The TACC saw the car in front starting to slow and slowed our car, but not fast enough to avoid closing on the lead car very quickly. In fact, quickly enough that the crash warning popped up on the display. I'd be watching all this closely and slammed on the brakes just as the warning came on and stopped in time without much problem.

IMHO, this is a flat out bug. TACC should brake hard enough to avoid collisions (Duh!). It had seen the car ahead slowing, why didn't it match it's deceleration rather than coming up to rear end the car?

I don't think this is a bug. Tesla flat out tells you that TACC has limited braking capability. The system is ultimately designed to assist a driver. So it is designed around the assumption that some scenarios will require a driver to step in and override it. If you approach it from the assumption that you shouldn't have to do anything you're going to be very disappointed with it. Despite all the hype, this is a step forward but it is not a self driving car.

I've found places that I really don't like using it. In the Bellevue, WA area merging from I-90 West bound onto I-405 north means you come onto I-405 in an exit lane and immediately need to merge left if you don't want off on the next exit. The overtaking mode tied to the left turn signal can make this a real pain because it tries to speed up into slowing traffic ahead of you. The other place I've found it to be really bad is coming into I-90 east from the 4th Avenue on ramp. After you pass the I-5 exits there is a sharp curve. There's not much traffic here but as you get around the curve you run into all sorts of traffic. TACC will try to accelerate up to your speed and won't see the traffic ahead due to the curve.

The vast majority of my high driving can be handled pretty effectively by TACC. I've grown more and more accustomed to letting it handle things that initially made me very uncomfortable. But I still realize I'm going to need to jump into situations that are unusual.

It's also very bad at dealing with people entering a highway who are a bit shaky on the concept of Yield. It doesn't notice them until they are almost completely in the lane, often getting much to close to the entering car. If the radar beam is that narrow, that's a bit broken too IMHO.

I don't think this is a matter of the beam being too narrow. You can tell this by the fact that cars leaving your lane don't get ignored by the radar until they're completely out of the lane. If the beam was too narrow then it'd accelerate after vehicles left the lane much quicker. Rather what you're seeing is probably an attempt to prevent the car getting false positives for vehicles in front of you from vehicles in neighboring lanes or stationary objects along the side of the road.

The driver can jump in and take over and deal with the partial lane change. But the driver and those following them will be much more surprised and much less capable of dealing with the problem of decelerations/stops for no apparent reason.

 

[RDoc]

If the TACC is designed to not safely stop in a normal traffic situation then Tesla is going to get itself in trouble IMHO. A boilerplate manual disclaimer is hardly sufficient.

In the event I described, the car was no where near maximum normal braking, never mind emergency level braking. I was easily able to stop the car safely. If the collision avoidance system means anything, the car must be capable of braking harder than that under computer control, so the TACC in a stop and go situation should be too.

If the design is that the car can't handle some normal braking situations, it should recognize that it's closing on a car or other obstacle faster than it's design braking limit and warn the driver very early that it can't handle it. Either that, or simply don't attempt to handle situations like stop and go driving that it can't deal with. Leaving the design such that it can handle 99% of the cases and silently fail on the rest is a recipe for some very bad consequences, particularly if that's the design philosophy Tesla intends on applying to their future "autopilot" mode software development.

 

[Stoneymonster]

If the TACC is designed to not safely stop in a normal traffic situation then Tesla is going to get itself in trouble IMHO. A boilerplate manual disclaimer is hardly sufficient.

In the event I described, the car was no where near maximum normal braking, never mind emergency level braking. I was easily able to stop the car safely. If the collision avoidance system means anything, the car must be capable of braking harder than that under computer control, so the TACC in a stop and go situation should be too.

If the design is that the car can't handle some normal braking situations, it should recognize that it's closing on a car or other obstacle faster than it's design braking limit and warn the driver very early that it can't handle it. Either that, or simply don't attempt to handle situations like stop and go driving that it can't deal with. Leaving the design such that it can handle 99% of the cases and silently fail on the rest is a recipe for some very bad consequences, particularly if that's the design philosophy Tesla intends on applying to their future "autopilot" mode software development.

There is NO Collision Avoidance System as I think you are describing.

 

[breser]

There is NO Collision Avoidance System as I think you are describing.

This is absolutely correct. Emergency braking is there only to decrease the energy of a collision (and the injuries/damage as a result). If you get really lucky it might be able to prevent a collision, but this is not the intent of the emergency braking.

TACC obviously wants to avoid collisions. But there are limits to the sensors and the algorithms processing the data. There will be scenarios where the car gets it wrong. That's precisely why Tesla calls these features Autopilot. They are driver aids, they help reduce the driver workload, they do not eliminate the need for the driver to be paying attention.

 

[S4WRXTTCS]

What I find bothersome from the entire Automotive industry (including Tesla) is the way these systems are presented versus what they're really intended to do.

All the MB commercials show some pretty amazing collision avoidance, but when you actually read the details on the system it reads just like the Tesla Manual does.

Even the name of some of the packages makes one really wonder what's being promised.

For example MB has Collision Prevention Assist Plus

Is that collision prevention?
Is it collision prevention assist? Where the plus is extra assist, and not sorta assist.

Luckily I got the 70D so I won't have to go around testing all these collision prevention/assistance/warnings systems to see which one did the best at simply not running into things. I know that's hard because I still haven't managed to not run into things while I'm walking, and I have lots of sensors (sight, sound. 6th sense, etc) along with a massive (although not so much anymore) neural net.

When I do get the 70D I do plan on testing it just to see what it's behavior is. I'll have to find one giant thing of foam and lots of aluminum foil.

 

[RDoc]

There is NO Collision Avoidance System as I think you are describing.

What I'm referring to is the "Automatic Emergency Braking — a new Collision Avoidance Assist feature" that supposedly can rapidly slow the car in an imminent collision situation. There was some question about how much braking the car can automatically apply, but if there's no mechanism for harder braking than I observed, then the system is completely useless. I doubt that.

I still think that the TACC software doesn't properly handle very ordinary traffic situations that it should be able to. In the situation I was in, the car sensed the slowing car in front, but didn't apply enough regen/braking to stop the car in time by itself. I was easily able to stop the car with heavy, but not emergency level braking, and at that point we were very close to the car in front.

If it's really true that the car can't stop itself any faster than what I saw, it should warn the driver a long time before it did that the TACC can't handle the situation so the driver knows to do something a lot earlier. As I said, the car clearly had sensed the car in front and was braking for some distance. If it couldn't stop in time, it should have let me know a lot earlier than it did when the collision warning came on just as I hit the brake. It had seconds during which it was braking, just not hard enough, to assess the situation. It looks to me like the algorithm isn't properly modeling the rate of closing and the amount of deceleration needed to stop in time. That's a pretty trivial algorithm, so I think they just got it wrong, i.e. it's a bug.

 

[brianman]

For example MB has Collision Prevention Assist Plus

1. Let's call it "Lifesaver."
2. Rookie. You can't make promises like that. Besides that's copyrighted. Also, you need to have a techno name like "Collision Immune."
3. Ok, junior. You did slightly better than the rookie. "Immune" lands you in court. "Collision Prevention."
4. We need to tack on "Assist" to be clear it's a tech that really doesn't replace the driver. "Collision Prevention Assist."
5. Next year we'll call it "Collision Prevention Assist Plus" because, like, it's better.
1.2.3.4.5. Good meeting. Let's go to lunch.