willow_hiller
Well-Known Member
I think there's still a place for the fan-videos, though. You need the over-enthusiastic videos to counter-balance people like Dan O'Dowd putting out videos of FSD Beta whacking dummies.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD Beta 10.69
- Thread starter Buckminster
- Start date
I think there's still a place for the fan-videos, though. You need the over-enthusiastic videos to counter-balance people like Dan O'Dowd putting out videos of FSD Beta whacking dummies.
some interesting turns here in heavy traffic. Beta did most of them well. It did fail a left turn and just go right at a different, slightly more irregular intersection later in the vid, but then did it well on the next attempt. Kind of a weird one at the end where it properly saw a gap and went, maneuvering around a car turning left from the main road.
The turn to the right had some sort of software bug. If you watch, you'll see that the car loses the beta visualizations right after that turn.
Other than that, a most impressive performance. I am very much looking forward to testing this software.
Other than that, a most impressive performance. I am very much looking forward to testing this software.
AlanSubie4Life
Efficiency Obsessed Member
Yep, looking forward to FSD doing this in future (where permitted by law).
I rate this as 5/7 on Chuck's turn, with 2 other left turns completed successfully, for a total of 7/9.
#5 -> too long a gap (6:34->6:40), should have gone on the over 5 second gap. However, some would call this a success. Waiting too long actually got other drivers to start going around the ego car, arguably. Obviously some of that behavior would have happened anyway, but the red car looks like it was no longer confident that the ego car was going to take the next good opportunity.
#6 -> 7:28. Poked the nose out, then stopped, forced driver to go from #2 (??? - it's unclear) to #3 lane (wasn't even in the target lane (I think?) but something about the movement was disconcerting I guess...I'm assuming it wasn't in lane #1 otherwise Chuck would have disengaged). Anyway. Can't be indecisive like this; this would certainly result in a disengagement. Some would argue this was a success, but I don't see it that way since it should not poke and stop like that. Just wait for the oncoming traffic to pass the "safe point" then go. There was a large gap afterwards. It's actually odd that it even budged; that was only a 4-5 second gap and it hasn't seemed to date like it would go if there was traffic in the near two lanes, which is fine (unclear about the third)...yet it did here.
Actually better than I expected. I guess we'll see if he posts more and we'll get a bigger sample. Would have to go back and review, but it's actually very close to 90% I think.
Last edited:
I suspect that the unexpected right hand turn was the result of the navigation crashing. FSD did it's default of making a right hand turn when it cannot continue straight at that point. A failure, for sure, but not a ULT failure.
AlanSubie4Life
Efficiency Obsessed Member
Yeah I didn't count that at all, it wasn't even clear a left turn was ok, but anyway it doesn't matter, I just ignored it. I counted the second attempt as a success so 1/1 on that left turn.
There was a later one where the assertiveness made a van move over but in my opinion the way that was done was ok. No hesitation and it didn't leave the median until it was clear and once it left the median it went.
Agreed granted I don’t expect this group to go full on optimistic lol. 6 seconds vs 5 seconds is apparently a fail unless someone crashes. I am rather impressed with the results from what I see. From the last version where it could not even Try this turn to full commitment says a lot about the teams commitment to getting results. Step by step wins the race.
AlanSubie4Life
Efficiency Obsessed Member
I'm interested in a driving feature that is usable. If it is going to wait on an obvious gap that is easy to fit into with zero hazard, I'm going to override with the accelerator and go. So that's a failure if it doesn't go. It's not a safety failure. It's just a usability failure.
There are obviously tons of improvements in capability demonstrated in what we see in 10.69. There's just no question that there are critically important capabilities now that didn't exist before. It remains to be seen how robust they are.
As you well know, I will definitely be positive about improvements once I get it, as I have been in the past. It looks to me like there will be significant improvements and I am most interested in how smooth it is. And I hope they tune up assertiveness as Elon said. It could be good!
The left turns are very important for demonstration of capability but equally important is just how well it goes in a straight line and stops. The stopping looks very problematic right now, but hopefully they'll improve that. On 10.12, I have to press the accelerator all the way to the stop line to make sure I get to the stop line as rapidly as would be possible without friction braking, and feather it just right to avoid jerking, and it looks like in 10.69 I'll have to do the same, but it sounds like they have added some capability that may make this behavior better (once assertiveness is increased). It just seems easy for them to feather in a max regen stop, to end right at the stop line, just like a human, and I'm not sure why they don't do it. In addition I have to press the accelerator to go and get back up to speed, at every single stop. So hopefully that will all be resolved soon.
He said "easily the biggest." But I'd agree with both based on what we've seen so far; it looks like the most significant improvement in capability we have seen in the nearly one year we've been doing this.
Last edited:
Less than 7%...but to your point, that still a LOT of dummies
jabloomf1230
Minister of Silly Walks
Chuck's heavy traffic video is impressive. The question is, how will this new paradigm translate to other unprotected intersections, especially roundabouts? Chuck tested a few more ULTs and other than the one software glitch, the car seemed to handle them adequately. If I had to guess, the new approach must have been in alpha for some time and Chuck's ULT posed a good option for publicly testing it. I originally thought Elon was taking a risk publicly targeting Chuck's ULT, but in retrospect, Elon probably already knew that 10.69 would be able to do the turn.
sleepydoc
Well-Known Member
Roundabouts aren’t unprotected. They’re a right turn with a yield sign.
Hopefully, we'll all have the opportunity to find out in a couple weeks. I already have a list of progressively more difficult ULTs and URTs just waiting for 10.69.x.
AlanSubie4Life
Efficiency Obsessed Member
It didn't seem risky at all to me. It was an easy turn, and with the correct capabilities it should be possible to make it. Remember that previously there was ample evidence that it could not "understand" occluded vehicles and would go at incorrect times (which led to disengagements).
Now the question is whether it can be done reliably and safely, now that it seems to have the basic capabilities and framework to be able to successfully proceed. Can it do it "100%" of the time as Elon said? Right now it clearly has issues but are they just going to be resolved with simple tuning, or are there issues with capability (not clear what that would be at the moment)?
Why the negativity? Fanboy or not you can still get some feedback by simply watching the video. Not everyone is as good as Chuck and Chris but that doesn't mean there isn't value. And nice to see Chris (DirtyTesla) has had zero phantom breaking.
James Locke has also published two videos and is very positive about .69. My sense is .69 will end up being considered the most improved version but only time will tell when Tesla pushes out future waves.
Yesterday I had my worst day ever with 10.13 so hoping to give .69 a try in the next 2-3 weeks.
Yeah, I also have a bunch of unprotected turns ready to test although very few are "Chuck Cook style" with a median crossover region between islands. There do seem to be quite a few general improvements in 10.69 for any style of turns that I hopefully will be able to experience soon, but around here, if FSD Beta could reuse the median waiting behavior for center turn lanes, that would probably help successfully navigate many more turns consistently.
I was hoping to see evidence of the car tracking vehicles through an occlusion. For example, you can still see vehicles disappear when there is another car briefly occluding them. Maybe the car is still tracking them internally, but the visualization is not showing them. They blink out then back in as they emerge from behind the occlusion.
This is important when the car is executing a maneuver like the one in Chuck's video where traffic was backed up and the car was threading through three lanes. If the car can project another car through an occlusion, it reduces the risk of getting T-boned.
willow_hiller
Well-Known Member
I'd recommend you watch Ashok Elluswamy's CVPR presentation, if you haven't already. But this tweet summarizes the part relevant to your comment:
The new volumetric occupancy network does seem to be designed to persist vehicles through occlusions. Hopefully the performance of those predictions will be improved in 10.69.1 and 10.69.2.
Negativity??? It is just a statement of fact and not meant as disparaging the videos or saying they don't have value. Just that they are on the subjective side and saying "best drive" is par for the course.
Also please let us know how you got 10.13?
AlanSubie4Life
Efficiency Obsessed Member
Right, but the observation here (I had not noticed but had not looked closely) is that visualized vehicles flicker when occluded in 10.69.
As was pointed out, maybe it still knows they are there for a while, not clear.
Need a lot more tests!
Similar threads
- Article
