Seems like FSD is a complete crock

[AlanSubie4Life]

I had a situation just today where a pick up truck passed me and veered a little bit over the middle line and my car on AP braked automatically. So yes, I have seen the cut in prediction work.

That's not prediction though. That's just detection of someone over a line. I have not yet seen any evidence of the car braking before another car enters the lane at all, which is what cut-in prediction would be.

 

[Dutchie]

Not true. You can teach computers to predict movement. Tesla already did it for predicting when cars will cut in font of you. The same technique could be applied to deer. Basically, you feed video clips into the computer and teach it to recognize the signs that the deer is about to run in front of you.

When on autopilot my car always brakes when another car crosses the lane further down the road, even if it doesn't need to as it will be long gone as soon as I arrive at the position where the car crossed the road. It is very bad anticipating these situations and uploading images does'nt do anything. It needs to upload video streams which it does not do.

 

[diplomat33]

That's not prediction though. That's just detection of someone over a line. I have not yet seen any evidence of the car braking before another car enters the lane at all, which is what cut-in prediction would be.

Except he wasn't completely over in my lane, his wheels just grazed the middle line. So I consider it a prediction.

 

[diplomat33]

When on autopilot my car always brakes when another car crosses the lane further down the road, even if it doesn't need to as it will be long gone as soon as I arrive at the position where the car crossed the road. It is very bad anticipating these situations and uploading images does'nt do anything. It needs to upload video streams which it does not do.

That's a different scenario that what I was describing in my earlier post. But yes, I have experienced your scenario as well. I am sure it will get fixed when Tesla implements cross traffic for "automatic city driving".

 

[Octo]

I think that’s the glass half full interpretation of phantom braking.

 

[AlanSubie4Life]

Except he wasn't completely over in my lane, his wheels just grazed the middle line. So I consider it a prediction.

Maybe. Obviously it knows the car is there so at some point it will brake. Response is probably also tuned to the situation...close proximity could result in braking. I am unconvinced that cut-in prediction has rolled out yet. Happy to hear/see more evidence though!

 

[diplomat33]

I am unconvinced that cut-in prediction has rolled out yet

I think you need to see more evidence.

Karpathy said during autonomy investor day that cut-in prediction had already rolled out to the fleet.

 

[diplomat33]

Here's a live blog of a new presentation by Tesla on their FSD computer with lots of technical info on how the FSD computer works.
Hot Chips 31 Live Blogs: Tesla Solution for Full Self Driving

 

[AlanSubie4Life]

I think you need to see more evidence.

Karpathy said during autonomy investor day that cut-in prediction had already rolled out to the fleet.

Yes, I do need to see more evidence. Would be great to see a bunch of videos of proper cut-in prediction - though it will be hard to see it of course.

The fact that it's apparently rolled out leaves me underwhelmed at its capabilities thus far. Presumably it will improve. When I see it slowing down in the correct situations (where spacing is adequate) to let someone in who starts signaling, that would impress me. Obviously cut-in prediction has to work without signaling too (arguably more important in that case), but the signaling case would be pretty cool to see.

Getting back to the topic, I'm not convinced based on what I've seen so far that Tesla is going to be able to make an easy transition to predicting the behavior of animals. I guess we'll have to see how things are doing when software that can take advantage of HW3 rolls out.

 

[diplomat33]

I guess we'll have to see how things are doing when software that can take advantage of HW3 rolls out.

The HW3 computer does look really good to me on paper. Of course, Tesla still needs to do the software to go with it. But I remain optimistic that we will see a big jump in Autopilot when Tesla starts fully utilizing the HW3 computer's capabilities. After all, we are still in the pre-FSD stage.

 

[EVNow]

That's not prediction though. That's just detection of someone over a line. I have not yet seen any evidence of the car braking before another car enters the lane at all, which is what cut-in prediction would be.

In my car (after some particular version that I don't remember, but I've posted here before) - on freeways the car doesn't break when someone cuts in / merges. It slowly slows down to get back to correct spacing. I'm on "chill" and 5 car lengths.

So, in this case, to figure out when it detects cut in - I watch for the darker grey changing from the car in front in my lane to the one cutting in. I've not seen many (any?) instances of the car recognizing cut in.

BUT, we may be misunderstanding. Perhaps earlier they were recognizing a new car in front much later ? That has changed with cut-in prediction, perhaps. We don't know.

ps : I think in many cases what they said in analyst day was misunderstood by the techie crowd to mean something else. A typical example is shadow mode - which could simply a way to explain what the trigger system they have implemented is, but many think of it to be much more full blown shadow mode that is always comparing what the AP would do to what the driver does and sends back differences.

pps : I think several things they showed were still in r&d stage (DoJo ?), not used in production. A good example is the 3D model they should that can be built using camera. They also showed how a path planner can be implemented by using NN.

 

[thewishmaster]

In my car (after some particular version that I don't remember, but I've posted here before) - on freeways the car doesn't break when someone cuts in / merges. It slowly slows down to get back to correct spacing. I'm on "chill" and 5 car lengths.

So, in this case, to figure out when it detects cut in - I watch for the darker grey changing from the car in front in my lane to the one cutting in. I've not seen many (any?) instances of the car recognizing cut in.

BUT, we may be misunderstanding. Perhaps earlier they were recognizing a new car in front much later ? That has changed with cut-in prediction, perhaps. We don't know.

Several times I've had my car brake to let in another car before I even realized the other car was considering merging, with and without blinkers. Once even a couple seconds before the other car merged (I was staring at the screen wondering why my car was suddenly slowing down and tracking a couple cars to the side). Also works quite well in traffic as most of the time the other car hasn't even left its lane yet and mine is already braking to make room. So I do believe there is some predictive black magic at play for certain situations.

 

[AlanSubie4Life]

Several times I've had my car brake to let in another car before I even realized the other car was considering merging, with and without blinkers. Once even a couple seconds before the other car merged (I was staring at the screen wondering why my car was suddenly slowing down and tracking a couple cars to the side). Also works quite well in traffic as most of the time the other car hasn't even left its lane yet and mine is already braking to make room. So I do believe there is some predictive black magic at play for certain situations.

I really will have to play with this some more. Never seen such behavior in my limited use. Every time someone changes lanes, it is apparently a complete surprise to AP.

 

[thewishmaster]

I really will have to play with this some more. Never seen such behavior in my limited use. Every time someone changes lanes, it is apparently a complete surprise to AP.

Some things do regress temporarily with some updates, so it might be working worse on newer versions right now (I’m still on 2019.20.x).

Another situation to watch for is when another car is making multiple lane changes at once - I’ve seen my car track others cutting across behind (potential obstacle while I am trying to change lanes on NoA, highlighted car on screen, NoA aborted the first lane change attempt) and in front (car cutting across but moving faster than me, also highlighted in gray, no need to brake).

 

[AlanSubie4Life]

All of the below with firmware 2019.28.3.1

One instance today, evidence of no actual "normal" cut-in anticipation:

Here is a completely failed-to-anticipate, obvious cut-in. This is my typical experience. You can see the nose of the car dip at around 20 seconds, when the blue hatch is finally fully in the lane. This was while on NoA. There was zero slowing until the car was fully in the lane. My car was in the process of tracking the prior vehicle (BMW X3?) (and was gradually increasing following distance since that vehicle had also cut in previously), when this cut-in occurred.

Shocker_Cut-in

Separately, I had two "Automatic Lane Change Cancelled" instances. Configuration: HW2.5. I'm not having NoA pick when to change on its own - I request the lane change then let the car do it itself. Car kind of approached the line, then swerved back, putting the error message on the screen (in the second instance it may not have displayed the message - when it refused to complete the maneuver, I overpowered it and took it out of Autosteer, but in the first, since there was no traffic, I just waited and tried again). I have theories about what happened in both instances, but it may also have been random.

If at first you don't succeed, try again...
AbortLaneChange
Cancelled autosteer with steering torque input upon refusal:
AbortedMerge