TACC Tuning?

[Saghost]

I agree, although in the interest of UI simplicity, I believe this goal can be achieved without requiring a separate mode. Implementing the items above should go a long way toward this. Everything can be summarized in one statement: sacrifice "perfect following distance" for smoothness of acceleration and braking, and everything will be nicer.

As I said in the initial post here, I think that the degree to which you'd rather have a smooth ride or a rigid following distance is implicit in the choice of following distance setting. A close following distance inherently has to follow the other car tightly, and I don't know why you'd choose a long distance if you wanted an exact match to the other car.
Walter

 

[Andyw2100]

I agree, although in the interest of UI simplicity, I believe this goal can be achieved without requiring a separate mode. Implementing the items above should go a long way toward this. Everything can be summarized in one statement: sacrifice "perfect following distance" for smoothness of acceleration and braking, and everything will be nicer.

I'd definitely have no problem with this as the first revision of what we have now, and personally I'd probably be happy with TACC if it never went beyond this.

But I expect there would be complaints from people who would want TACC to drive more aggressively, so I'm just trying to cover all the bases. Those people could make the same argument that I've made in other threads where some people are saying that you should never warm the battery when not on shore power. My argument, which is what the people who want the more aggressive TACC could argue, has been that if you know you don't need to worry about range, and prefer some benefit that costs some efficiency, there's no reason not to choose the benefit over the efficiency. In the battery warming discussion that benefit is regenerative braking at the start of the trip. In the TACC discussion that would be more aggressive acceleration. I think in general the argument that we should be able to choose to give up efficiency for a benefit that we would like to take advantage is a valid one. So I'm just trying to plan ahead for the people who will make that argument.

For now, as you suggest for simplicity's sake, the way you summarize things above is just fine, and would be a great start.

 

[Brunel]

Just to bump this conversation with a specific question, prompted by another thread where a Model S rear ended another car and motorcyclists started commenting about safe distances.

Have people experienced TACC with motorcycles in front and if it only works in a central sweet spot? I see bikes swerving in and out on the motorways and keeping close to the next lane divider to slip between cars, so if TACC did not pick up the bike the car could follow dangerously close.

 

[billarnett]

I followed a motorcycle on the freeway today using TACC. It worked exactly like it does following a car. The bike was riding normally in one of the tracks, though, not lane splitting.

 

[Todd Burch]

I haven't noticed anything unusual using TACC and following motorcycles. Behaves just as with cars.

 

[Vger]

Those of you who've been using TACC these last couple weeks: Do you have a feel for how precisely it mimics the acceleration/deceleration of the car in front of you? Does it vary when set to different distances?
...
Hence the question: Is Tesla already doing this? If not, do you agree with me that they should?
Walter

Great question, well posed. I agree with the combined function idea. Too many options is going to get rapidly unmanageable with this car. Designing coherent scenarios, with co-varying parameters, makes good sense to me, in this particular case, but perhaps others going forward.

Actually, now that I think about it, they are already doing this type of thing with Range Mode— combining HVAC and power train efficiency optimizations.

 

[mikeash]

As I said in the initial post here, I think that the degree to which you'd rather have a smooth ride or a rigid following distance is implicit in the choice of following distance setting. A close following distance inherently has to follow the other car tightly, and I don't know why you'd choose a long distance if you wanted an exact match to the other car.

It's not quite implicit for me currently. I generally leave the following distance set to 7 (for the few days I've had the car!) and it's still more aggressive than I like at accelerating once the car in front of me starts to move, or changes lanes. It's not too bad, but I would like to dial it back by 1/3rd or so if I could. I love the acceleration the car can do, but cruise control on the highway isn't really the right time for it.

 

[TTT]

Yesterday I drove up to Culver City SC and back from San Diego almost exclusively using TACC and drove around San Diego today using TACC. For the most part it was a great drive but I think there needs to be more refinements before it's ready for autopilot. I hope Tesla is downloading everyone's TACC performance, analyzing and incorporate into the next update.
1. Stopping needs to be more gradual, especially in traffic, the camera should pick up the brakelights ahead and start slowing down. With TACC engaged at 75mph, and four lanes of brakelights dead ahead, TACC stayed at 75mph while cars on either side of me slowed down. It wasn't until I was uncomfortably close before TACC started to decelerate.
2. As California drivers are, they tend to cut in front of you when you have an ample safety zone between you and the car ahead, TACC doesn't cope with that type of traffic very well, three times my collision display and chimes sounded as cars cut me off and didn't seem to brake soon enough.
3. TACC should also calculate the appropriate speed in turns and adjust accordingly. It's not a problem if you have a car in front of you in the turn, but if there isn't, TACC maintains the set speed even if it's too fast for the curve in the road.
4. I was cruising down the diamond lane right behind another MS and it dawned on me that it would be cool if the cars would sync up together so the following cars could share the lead car's sensors so that response time reacting to leading non-Teslas can be shortened.

 

[Andyw2100]

I've recently noticed one aspect of TACC that definitely seems to be working better now than it was in earlier versions of the firmware, though I can't pinpoint when the change took place. What I've noticed is that if the TACC has no target car, and a faster car passes you and then cuts in front of you, cutting into the following distance you have set, the TACC is much less likely to actually decrease your speed. There are still variables, like how much faster the other car is travelling, and how close to you it was when it merged back into your lane. But in general, TACC is handling these situations much better.

 

[Irish Dan T]

I'm not sure if it was an update or something my SC did on a recent visit, but my #6 follow distance is now the old #7. The new #7 gives more distance to the followed vehicle than ever before and therefore more time to react at higher speeds. I also get a smoother acceleration on resume. I like it. Anyone else notice this? There's probably a thread somewhere that I couldn't find on the subject.

 

[spentan]

I'm not sure if it was an update or something my SC did on a recent visit, but my #6 follow distance is now the old #7. The new #7 gives more distance to the followed vehicle than ever before and therefore more time to react at higher speeds. I also get a smoother acceleration on resume. I like it. Anyone else notice this? There's probably a thread somewhere that I couldn't find on the subject.

What firmware version are you running?

 

[Cyberax]

If the forward looking Radar that is tracking the vehicle in front is a Doppler Radar then the algorithm could use the Doppler info as a direct measurement of relative speed (closing speed), in addition to distance, to greatly improve performance and increase efficiency.

It's a Doppler radar, but it has no spatial resolution to speak of. So tracking a moving car in front is easy (since it returns a nice radar return with a clear relative speed) but detecting stationary vehicles is a problem.

 

[garygid]

If the radar can pick out a good return with you going 1 mph faster than the car ahead,
or 30 miles faster, a good return with you going at any speed should not be ignored.

Cars ahead going slowly are detected, but stationary objects (a high relative speed
essentially equal to my car's speed) are ignored? That sounds like it needs to be
"fixed". Approaching a wall, an overturned truck, stopped traffic, very slow traffic,
all are important.

Perhaps the original software was just written to ignore stationary objects.

 

[RDoc]

It's a Doppler radar, but it has no spatial resolution to speak of. So tracking a moving car in front is easy (since it returns a nice radar return with a clear relative speed) but detecting stationary vehicles is a problem.

I'm very doubtful of that, do you have any source for this assertion?

First, Doppler radar detects and measures distances to objects the same way non-Doppler radar does, it just adds the capability to sense relative velocity. Second, if your car is moving and the other car is stopped there still is a relative velocity, it's just negative, it's no different than if the other car is moving slower than you. Even if you are both stopped, there's still a relative velocity, it's just zero, which is just as respectable a number as any other.

 

[beeeerock]

Cars ahead going slowly are detected, but stationary objects (a high relative speed
essentially equal to my car's speed) are ignored? That sounds like it needs to be
"fixed". Approaching a wall, an overturned truck, stopped traffic, very slow traffic,
all are important.

Perhaps the original software was just written to ignore stationary objects.

I haven't seen this in my experience. In fact, just the opposite. There have been at least three occasions where (and this seems to be the pattern) I've come up a hill that crests and breaks to the left. There has been a car parked on the right shoulder of the road. TACC has panicked that I was going to hit the parked car and piled on the brakes... then realized the reality of the situation and released them.

The occurrences were all in different locations, but the basic conditions were the same. I suspect the crest curves had too low a K value (rate of change of grade too quick - 'sharp' vertical curve) and that made the road paint lines invisible or difficult for the camera to resolve. The car was therefore right in my path as far as TACC was concerned. Short of putting the camera on a pole above the car, I don't know how this could be corrected... but it suggests to me that TACC is very much aware of stationary objects. It wakes me up, I'll tell you... and the person behind me too... :redface: However I'm aware of this behavior now and knock the TACC out of operation when I think it's going falsely trigger.

Here's a good one that almost got me yesterday... coming up a hill with a passing lane. Two trucks trying to swap order at the top, where the passing lane ended. Slow one right over on the shoulder, faster one going wide into the oncoming lane. TACC didn't see *either* truck and was looking right up the middle between them... I had to hit the brakes to avoid becoming a Tesla-meat sandwich between the two of them. I could have swerved left or right to give TACC a better look at them, but it seemed more prudent to just hit the brakes... LOL

On a positive note, the camera appeared to read 'Thank You Resume Speed' as I left a construction zone. I'll have to see if I can reproduce that one... certainly it's hit and miss on the big speed signs themselves... there's a divided highway posted at 120 km/h that it won't read ANY signs on. Seems to default to the older 110 km/h that likely lives in the Google database. But for some reason, it sees a section at 80, then another just a km further at *60* km/h. It would NOT be good if the car was able to follow those speed signs autonomously...! But that's off the TACC topic.

 

[Cyberax]

If the radar can pick out a good return with you going 1 mph faster than the car ahead,
or 30 miles faster, a good return with you going at any speed should not be ignored.

It's more complicated.

Cars ahead going slowly are detected, but stationary objects (a high relative speed
essentially equal to my car's speed) are ignored?

Yes. Essentially, the radar "sees" a lot of reflections from everything around it.

Imagine it like this - a radar emits a pulse and immediately starts receiving reflections from the road surface, nearby walls, other cars. A car directly in front of it will cause an especially powerful reflection - but a concrete median can also produce a strong radar echo.

However, if an object _moves_ relative to the background then it's very easy to detect it - its radar return will be Doppler-shifted differently than that of the objects around it. It's like an automatic filter.

- - - Updated - - -

I'm very doubtful of that, do you have any source for this assertion?

Yes. Tesla uses Freescale MR2001 radar transceiver - you can check its datasheet here: Multi-channel 77GHz Radar Transceiver Chipset|Freescale

First, Doppler radar detects and measures distances to objects the same way non-Doppler radar does, it just adds the capability to sense relative velocity. Second, if your car is moving and the other car is stopped there still is a relative velocity, it's just negative, it's no different than if the other car is moving slower than you. Even if you are both stopped, there's still a relative velocity, it's just zero, which is just as respectable a number as any other.

Doppler shift allows to identify objects that move relative to the background objects (yes, background objects are also Doppler-shifted, but they are Doppler-shifted _identically_).

- - - Updated - - -

First, Doppler radar detects and measures distances to objects the same way non-Doppler radar does, it just adds the capability to sense relative velocity.

BTW, Tesla's radar uses a modulated constant beam - it constantly "illuminates" the surroundings with a beam modulated by pseudo-random numbers. The modulation is frequency-independent (to avoid problems with Doppler shifts) and the receiver does the reverse - it finds the match for the incoming de-modulated bit stream and also extracts the frequency shift.

That's also why it's not a problem if more than one Tesla is on the road - their radars will use different bitstreams to modulate the beams and receivers will simply discard the incorrect radar returns.

 

[beeeerock]

Yes. Essentially, the radar "sees" a lot of reflections from everything around it.

Imagine it like this - a radar emits a pulse and immediately starts receiving reflections from the road surface, nearby walls, other cars. A car directly in front of it will cause an especially powerful reflection - but a concrete median can also produce a strong radar echo.

However, if an object _moves_ relative to the background then it's very easy to detect it - its radar return will be Doppler-shifted differently than that of the objects around it. It's like an automatic filter.

So in my experiences with crest curves and the road breaking left at the top, presumably the camera provided the information needed to cause TACC to hit the brakes hard for a moment? Presumably there is more than just the radar returns used to run TACC?

 

[Cyberax]

Possible. I won't be surprised if the radar is actually used only as a last-resort backup to detect impending collisions and camera does the bulk of heavy lifting.

 

[RDoc]

There seems to be some misunderstanding of how the TACC and more generally continuous wave radar works. To start with, it's not like a speed radar that can only measure velocity. The radar transmitter typically emits a stream of "chirps" which typically start at a low frequency and rapidly increase frequency at a known rate (there are various profiles, but apparently a simple sawtooth is typical), then the receiver looks at the echo and matches the previously sent chirp profile to the received one. Because of the delay in getting the return due to the distance out and back to the target, the chirp is delayed, but still has recognizably (by the electronics) the same shape. By measuring the time delay of a feature on the returning chirp the distance can be computed. Relative speed can be extracted by buffering a bunch of these returns and essentially looking at how the chirp spacing is changing.

Bottom line is that both accurate relative speed, including zero, and distance are available, and because the beam is electronically steerable, it can get target direction information at the same time.

 

[beeeerock]

Bottom line is that both accurate relative speed, including zero, and distance are available, and because the beam is electronically steerable, it can get target direction information at the same time.

Whoah! That's very interesting! Any idea how this is done? Is the transmitter horn/dish actually moving back and forth or is there some sort of arrangement of antennas utilizing diversity in some form? You say 'electronically steerable' so I doubt there is anything physically moving.

I saw something today that makes me wonder about this. I caught up to a black dually Chev in some corners along a lake, against a rock cut. I saw the TACC icon lock on and my speed started to reduce. Then the truck drove into a dense shadow and it sort of got lost for a moment against the dark rock, the truck being black and all. The TACC icon disappeared for a moment and the car surged ahead... then the lock was regained and I slowed down again. This all happened in the space of a couple of seconds.

The corners aren't tight enough to lose coverage of the vehicle ahead - at least, it hasn't happened before and I drive this highway regularly. It made me wonder if the camera had been involved with the radar and hadn't adjusted exposure for the shadow quickly enough to retain its visual lock (like my eyes!). However, if the radar is sensitive to direction, perhaps there was some confusion about which way I was going relative to the truck... having only just locked on before unlocking.

It's super-interesting stuff and I wish the average Joe (or pedantic SOB's like myself) could be given access to the finer details of the programming logic (not the code itself), because these sorts of clunks are probably quite explainable with better understanding of the system.

- - - Updated - - -

Possible. I won't be surprised if the radar is actually used only as a last-resort backup to detect impending collisions and camera does the bulk of heavy lifting.

If the radar is steerable and the camera is able to somewhat match directional distance information with its optical information, it would seem the radar is used for depth perception. I'm a little surprised that the radar transmitter isn't up in the windshield beside the camera... vertical crest curves (or sag curves) could make thinks less reliable with the lower position.