FSD never merges into highway from on-ramps

[70Dan]

FSD will never merge onto a highway from the on-ramp unless I signal it to do so.

If I don’t signal, it will wait until the very end of the on-ramp to merge, all the while not trying to adjust its speed to find a hole when there is some traffic. In fact it gets itself into trouble more often than not because it will immediately speed up once the car in front merges, thereby overtaking it from the right, often without enough room to pass before the end of the ramp, instead of just merging behind.

Other than that my FSD behaves normally. Many long, otherwise perfect drives have to be interrupted because of this bug (or missing feature).

Is this just a problem with my car?

HW3 Model 3 in the Ottawa region in Canada. On 2023.27.12.

This is the most frustrating issue for me after the auto wipers being worse than useless; they are downright actively dangerous in winter whenever they suddenly decide to spread road grime all over the windshield, instantly making it opaque and impossible to see through. HEY TESLA, YOU CANNOT DRY WIPE ROAD GRIME. Stop trying to kill me.

 

[Tam]

FSD will never merge onto a highway from the on-ramp unless I signal it to do so.

If I don’t signal, it will wait until the very end of the on-ramp to merge, all the while not trying to adjust its speed to find a hole when there is some traffic. In fact it gets itself into trouble more often than not because it will immediately speed up once the car in front merges, thereby overtaking it from the right, often without enough room to pass before the end of the ramp, instead of just merging behind.

Is this just me?
HW3 Model 3 in the Ottawa region in Canada. On 2023.27.12.

This is the most frustrating issue for me after the auto wipers being downright dangerous in winter.

Other than that my FSD behaves normally.

Either a human programmer or an AI needs to write up those code lines for your scenario.

 

[mgs333]

Either a human programmer or an AI needs to write up those code lines for your scenario.

It's not like that's some unusual, rare scenario. I run into the same situation as the OP every single time my car is merging so that I no longer let the car merge. From what I've heard others on here say, this seems like it's by design (or lack thereof).

 

[Tam]

It's not like that's some unusual, rare scenario. I run into the same situation as the OP every single time my car is merging so that I no longer let the car merge. From what I've heard others on here say, this seems like it's by design (or lack thereof).

Yes. If there are no programing codes for Autopilot to stop at a red light but there's one for FSD to do that, then it's a design for cheaper Autopilot to lack that traffic light response.

That's the same with merging competency: someone (human) or something (AI) needs to design or write the codes.

Elon Musk Thinks Tesla Made a Huge AI Breakthrough. Is the High-Flying Stock a Buy on the Dip? | The Motley Fool

Tesla just might have made progress toward achieving what some call the "holy grail" of AI.

www.fool.com

The same principle is applicable to Waymo that can competently merge because of their software team who spend time to write up the algorithm.

Tesla can do the same too but it just doesn't have the Waymo team with them.

 

[Todd Burch]

It’s not your car, nor a HW3 vs HW4 problem. That’s just how it behaves right now.

I’m expecting v12 to fix this, but we’ll see.

 

[SidetrackedSue]

In Driver's Ed, I was taught the driver joining the highway is supposed to accelerate to the highway speed, then merge at the end of the on-ramp or when it is safe to do once the car has come to the highway speed.

And this seems to be what FSD is doing.

Why this doesn't work/feel right is that the car is hell-bent on getting to the speed of the highway, not the speed of the traffic on the highway. So the merges at the end are relatively terrifying, even if they are successful. In other cases, the freeway on-ramp becomes the exit ramp for the next exit. FSD has been known to not try to merge in this case, apparently not recognizing the ramp has become an off ramp and ends up on the shoulder beyond the exit point. I don't use FSD for on-off highways so have no idea if this known problem has been fixed.

A good driver, brings the car up to the speed of the traffic in the next lane, then looks for a chance to merge. If there's no obvious, smooth, place to merge, then they go to the end of the on-ramp and zipper merge.

In my case, at my freeway on-ramp, there is a semi-permanent lane closure of the lane I'm merging into about a KM away and so almost every time I come onto the highway at my on-ramp, the traffic is traveling at about 30 - 50 kph instead of the 100kph posted because the traffic is moving from 3 lanes to 2 and this choke point is slow in times of heavy traffic or accidents, of which there is one almost daily. I'm trying to merge into a solid line of traffic with no gaps to merge into (and distracted drivers, themselves trying to merge to the lane to their left.) So I will go to the end of the on-ramp at just above the speed of the traffic (but still a fraction of the 100kph) and zipper merge at the end, and immediately prepare for the second zipper merge within a short distance. There is no friggin way I'd trust FSD in this situation.

 

[RTPEV]

I've found it to actually be better now than it used to be. It seems to do just fine now when there is little to no traffic to merge into.

I don't know how it does in heavy traffic because frankly I don't trust/dare to let it do it itself in those situations.

As for whether there needs to be an "algorithm" written for this, this is certainly not the direction that Tesla is going in. Their approach is to develop a general AI solution where path decisions are made entirely by the neural net based on training. It's not that I don't think that approach can ultimately be successful, but we probably shouldn't hold our breath for it to be able to handle every situation it may encounter.

 

[70Dan]

Thanks for the feedback. I was almost in disbelief that it could behave like this on purpose since that seems like a pretty important feature.

But then roundabouts are also very scary at times and nowhere near production ready.

Can’t wait for end to end neural networks in v12.

On a side note, if someone manages to retrofit a rain sensor to fix auto wipers, they might become rich selling the product.

 

[Goose66]

I’m expecting v12 to fix this, but we’ll see

That’s a pretty noob response for a 13 year member!

 

[Todd Burch]

That’s a pretty noob response for a 13 year member!

I choose to live life on the optimistic side.

 

[Goose66]

Just to say, from a practical standpoint, if they didn’t bother to code it in the first 11 versions, I doubt it will be a priority for gathering and curating training data in v12.