FSD Beta Videos (and questions for FSD Beta drivers)

[GlmnAlyAirCar]

What I enjoy is hitting the speed bump on my e-bike at 25mph (in a 25mph zone), and then looking back at the motorist that has to slow down for the speed bump. Sure eventually they catch back up, but for that one moment I'm beating the car.

In any case speed bumps are something that humans both gauge, and they learn from.

It's also dependent on what I'm carrying with me. If I'm in my Jeep I don't usually have to slow down too much unless my passenger has coffee, and I like the passenger.

It's also weight dependent.

Back when I had my 2015 Model S I accidently scrapped on a speed bump I had no issues with at other times. Why? Because I had 4 adults with me in the car.

Speed Bumps, Pot Holes, etc are all things the car is going to have to contend with. Unless they want to give free CyberTrucks to every FSD owner. That's fine by me too.

Of course I think people are using different definitions of speed bumps. Most of what I have seen in the FSD videos are what would be called “speed humps“ around here - wide humps that can be taken at 20 mph. What we here in the northeast call speed bumps are narrow, maybe 16“ wide, bumps that only one axle can go over at a time. Even with good ground clearance you have to move pretty slow to avoid high bandwidth impulse (I.e., sudden, jarring) up and down motion . The lazy speed “humps“ don’t have this jarring effect.

so I may be using different language than others in this thread.

 

[S4WRXTTCS]

Of course I think people are using different definitions of speed bumps. Most of what I have seen in the FSD videos are what would be called “speed humps“ around here - wide humps that can be taken at 20 mph. What we here in the northeast call speed bumps are narrow, maybe 16“ wide, bumps that only one axle can go over at a time. Even with good ground clearance you have to move pretty slow to avoid high bandwidth impulse (I.e., sudden, jarring) up and down motion . The lazy speed “humps“ don’t have this jarring effect.

so I may be using different language than others in this thread.

Regardless of definition there are numerous types, and the car needs someway of figuring it out.

There is one speed bump in a hump in a parking lot near me that I've never attempted to take the Tesla over. There are all kinds of marks on it from other vehicles, and I have no idea why the property management has this crazy combo thing.

I look forwards to seeing if FSD beta will even attempt going over it.

 

[scottf200]

Good turn. Unbelievable graphics.

 

[powertoold]

Wide-release fsd beta should also do well in stress-test conditions, but here’s AI addict doing a routine drive:

 

[msanghavi]

Good turn. Unbelievable graphics.

I'm confused why it did a good job here, but failed in other beta 9.0 runs. Could it be because of maps? @diplomat33

 

[diplomat33]

I'm confused why it did a good job here, but failed in other beta 9.0 runs. Could it be because of maps? @diplomat33

It is hard to say. We would really need to look at similar left turns where V9 fails and try to analyze why it failed in those examples. Also, Tesla does not use HD maps. Tesla does use regular maps. Maybe in the cases where V9 fails, there was a problem with the map. That is possible. But Tesla relies on camera vision for perception. So it is also possible that in the cases where V9 failed, there was a problem with the camera vision. But there could be a lots of things that could cause the camera vision to make a mistake. Even the same intersection, the traffic could be different, the lighting could be difficult. There are many factors at play that could affect V9. So it is difficult to know what caused the failure.

 

[mhan00]

Wide-release fsd beta should also do well in stress-test conditions, but here’s AI addict doing a routine drive:

That first unprotected right was terrible. Took waaaaaay too long and it looked like it had to creep way too far out before it finally went. Good thing the road was basically empty.

4:00 - car gets to Home Depot no issues, but they cut the vid right before it looked like the car was starting to go straight for an island in the parking lot and then resumed a couple of seconds later. Maybe it was nothing but the camera cutting out for some reason, but I’m guessing there was an intervention which is okay since the Beta isn’t supposed to do parking lots yet, but would still be a little alarming. It didn’t look like the islands were shown very well on the screen so not sure how well the car saw them.

5:30 - car does a solid job driving through a parking lot and navigating to the road, but on a turn goes a little wide and an oncoming car in the parking lot had to swing wide around the car. Passenger comments that the driver gave them a look and the driver didn’t intervene since he didn’t feel there was any danger of collision since the car saw them and there weren’t any parked cars preventing the other driver from maneuvering around them, but still not a good job. But again, Beta isn’t really meant to be used in parking lots yet so not a big mark against it, just a something to note.

The unprotected right after that is slow and overly cautious and the driver hit the go pedal to make the car go, but I think that’s good for now. Better to err on the side of making people a little impatient than risk safety, which is what I’ve seen in some vids of other autonomous systems too. Car didn’t take a fairly large gap and then didn’t go when two cars after that were both signaling that they were entering the parking lot. Driver mentioned that most people would go when they see a signal, but I’ve seen too many people who clearly forgot they had their turn signal on or who turn on their signal too early or by mistake to take that as a real indication it’s okay to go. I always wait for additional signs like slowing down and seeing the car start to turn before I personally go myself. Especially during the beta, it’s better for it to sit there forever and require the driver (who should be attentive and ready to take over anytime already) to hit the go pedal than to risk a collision. Obviously that needs to be fixed eventually for true FSD, but while they’re refining it, good job.

Overall a good drive that shows the system does pretty well in easy environments like most suburban areas are.

 

[mhan00]

Something that I think might be helpful for Beta drivers is for Tesla to put in flashing red waves similar to the blue AP nags when it’s at an unprotected turn/straight, but on the right and left sides to show that the Tesla sees the oncoming traffic from each respective side so the driver knows it won‘t try to just go. That way the driver can see the flashing red on the screen in their peripheral vision while they’re checking for traffic.

 

[Chazman92]

FSDBeta 9.0 - 2021.4.18.12 - Comparison drive in preparation for v9.1 - Memorial Park Route

 

[diplomat33]

FSDBeta 9.0 - 2021.4.18.12 - Comparison drive in preparation for v9.1 - Memorial Park Route

I really appreciate you posting these videos. Thank you!

I hate to be the pessimist, but I am still noticing a lot of basic issues. You had a bunch of disengagements in that video that seemed quite basic to me. IMO, the perception seems pretty good overall. But it looks like the path planning needs a lot of work.

 

[cbrigante2]

I really appreciate you posting these videos. Thank you!

I hate to be the pessimist, but I am still noticing a lot of basic issues. You had a bunch of disengagements in that video that seemed quite basic to me. IMO, the perception seems pretty good overall. But it looks like the path planning needs a lot of work.

I think that is what makes it a good baseline video to compare 9.1 with. It's really our first chance to see how fast these issues are going to be improved.

 

[n.one.one]

There should be 9.1 videos in the next few hours.

 

[powertoold]

Significantly smoother (visuals) and significantly more confident than V9 IMO (decision making, especially on the unprotected left after speed bump stop at 3:52):

 

[Matias]

That was nice. Though route was easy and traffic was very quiet. But nevertheless.

 

[powertoold]

That was nice. Though route was easy and traffic was very quiet. But nevertheless.

Yup, but remember that FSD doesn't take those things into consideration. As humans, we do. There are thousands of potential false positives that V9.1 doesn't get caught up in. The drive seems simple, but it's amazing that it's doing it purely by vision without constantly mistaking this and that for cars / cones / pets / peds / etc. etc.

Edit: More here. The improvement in 3 weeks is astonishing. Looks like we'll get wide release within 2 months at this rate of progress:

https://twitter.com/x/status/1421396399107702787

https://twitter.com/x/status/1421410678288519170

It's very clear at this point Tesla will achieve fsd at human-level+ statistics soon (at wide release).

 

[qdeathstar]

Let’s wait and see what ppl critical of fsd have to say....

 

[qdeathstar]

My lack of understanding of American idioms and culture is evident. I don’t understand this response.

Is it meant to say that i’m childish or grumpy old man or something else?

I don’t understand I don’t understand I don’t understand

 

[Mengy]

Very anxious to see the V9.1 videos today just to see what the level of progress is!

 

[Bladerskb]

Edit: More here. The improvement in 3 weeks is astonishing. Looks like we'll get wide release within 2 months at this rate of progress:

It's very clear at this point Tesla will achieve fsd at human-level+ statistics soon (at wide release).

lmao give it a rest. Your ignorance is astonishing.

 

[Terminator857]

.... The improvement in 3 weeks is astonishing. Looks like we'll get wide release within 2 months at this rate of progress:

Agree the amount of improvement is significant. The lines aren't bouncing all over the place. Perhaps that makes the videos seem impressive. Color also seems improved.