I have found 12.5. to be great. Feels more and more like a real human driving the car. The rate of progress is great.

As I understand the system it's not truly an end to end black box system as they have said, they also feed it what you are seeing on the visualization which is road perception, some obstruction detection, maybe the occupancy network too and more. Those things might detect the boom. The end to end part only learns what's implied is the training set so it doesn't have any generalized detection limited to any particular height but limited to examples it has seen with less than perfect recall. So if a boom looks different or is placed higher than anything it has ever seen it might not see it at all. There is a lot of fuzziness when it has to learn everything on its own.

Why didn’t you report the failure to stop? All you have to do is hold the voice button, say bug report, and a one sentence description? You will get much faster fixes to unique situations like this and you will accelerate the March of 9s to autonomy rather than waiting for the system to learn.

The release notes are no longer ‘granular’ because they no longer control the changes, so getting definitive numbers is harder. It’s just amount of data, quality of data and car compute optimization. They’ve given up micro control to A.I.

Just curious on your overall thoughts of the extender? I have that exact camera and lens but contemplating on the extender.

Hi. First time viewer. Thanks for your effort. I would be curious to know what camera you are using and how it is mounted. Thanks. Be safe.

Regardless of whether people think it would have suddenly stopped, it is not safe human like behaviour.

this is why they don't want Americans to own firearms !!!!!!!!!!!!!!!!!!!!!!!!!

They are celebrating Maduro's victory. Why they are even here? Aren't they supposed to be fleeing Maduro's tyranny?

I’m also running 12.5.1 now and noticed that the car no longer tries to kill me by actually yielding to traffic in a sharp curve merge instead of taking it on at full speed. Speed also seems improved over 12.5, where before in that version, the speeds the car would travel at would be under the speed limit often.

As a human that’s never seen that gate or barricade before, I would have done the same thing. Stopped at the first gate, and if there’s a safe way to get around it, take that path, but obviously stop at the gate if it doesn’t open. I wouldn’t know to back up to get the bar code to read again.

How on earth would an AI system be trained on barricades. It can get out of parking lots with baracades either. Use ur own brain please. FSD can work miracles dude.

These are the training challenges left. Things such as level crossings are at best a toss up if FSD 12.5 will pass. Level crossings are deadly and if FSD gets stuck because it decides to drive through on when there is no clear exit then drives lives are at high risk.

I am going to start muting channels looking for engagement with cheap drama. Yours might be a good start.

Use your voice to provide feedback to Tesla with these disengagement interventions when you press the break during FSD (Supervised). The more you provide feedback the better Tesla's will get at these situations. I don't own a Tesla but I've been following along FSD for years. At this point when I do own a Tesla in the future I'd be right at home in the driver's seat.

They just need to make those barricades more visible. Easy to hang some high visibility material from them.

the private interfaces between public roads and private ground will pose a major problem for fsd, because they are all different.

What are they celebrating? I noticed a bunch of Venezuela flags but not one American flag

This can be a huge problem that needs immediate attention: My friend and I received 12.5 last week and we are very excited about how well it is doing. But, this is a comment he sent to me "On Palm Ave heading to Monterey hwy. Train crossing arms came down while I'm a good distance away, with full lights flashing. FSD was focused on traffic light and ignored train tracks. I jumped on brakes to stop inches from crossing arm just as Caltrain roared by." I don't know if 12.5.1 has fixed this yet.

Good tests, thanks!

Highly recommend auto speed as it’s clearly the future of FSD. It doesn’t always go the same speed as I would, but it eliminates a huge number of ancient speed related issues associated with map issues.

Humans have two eyes on the front of the head, like most predators, to give good range finding. Perhaps that helps with gates. Tesla vision stitched together from widely spaced cameras pointed different directions seems more like cow vision (or other prey), with eyes on the side of the head to give a wide field of view, but i suspect less good for tracking, like you'll often see a human do with their eyes and head when watching something, or a cat hunting. Watch cows or deer, and they seem less good at this.

the barricade is really a rare and low priority case for FSD. There are quite more high priority things(pothole, road sign, speed control etc) that needs to be addressed first. It is pointless to keep testing barricades at the moment, as it won't be fixed for sure anytime soon.

Thanks for the video it will be interesting to see how long before they fix this ,it is a unique/hard problem

What hardware version are you running (3 or 4)?

Thanks. We have Chuck’s turn and now Don’s Gates. Very good ways to benchmark progress.

Blind to barricades yet it stopped?

There is a video with "Dr know it all" that discuss this issue. Known since first FSD version. The vision system is unable to see this horizontal structure for what it is. It could be a curb 50 meters ahead or something. The problem with vision and no LIDAR or radar is that everything gets really really really hard, and also conflicting with other stuff to see. Imagine being a soldier and you see another soldier. But you are color blind and cannot see the uniform color. You have to see if he shoots at you to know what to do. Then you have to start looking for other queues to discriminate. The car picks up the pixels, for sure but as know it all suspect there may be a horisontal scan line issue from the harware. But more likely i think Tesla runs into a conflict here in the machine learning. If it would be good at seeing this it will loose ability to see/interpret other stuff, like curbs or road markings, like a stop line or speed bump line. This is why the idea that Elon claim that humans have only vision and can see and drive is flawed. When have humanity and nature not gone for super human abilities, bats, dolphins, nigth vision goggles etcetera. So what happens here when the gate lifts and the car slows is that the bar goes diagonal and the vision system immediately regonizes it as an obstacle. So if Tesla could by machine learning get the car to recognize the stand for the bar and othe queues indicating a gate environment it will be able to see that the bar is a bar, probaly. There is a hard and an easy way to do stuff. Us engineers lear to always use the least compilcated approach required. Elon is not an engineer.

Nice video thank you

Focus on testing what FSD has been trained to do. It is interesting to see what the car will do in random scenarios but it is meaningless.

When it switches from the new stack to the old stack when entering and exiting the freeway It can act very odd. Hopefully this will be fixed when they merge the end to end stack with the highway stack.

That is similar to parking case when you need to enter and exit with barricades. I don’t believe this version (or any prior versions) does that yet. It likely can’t handle emergency vehicles yet either, so I wouldn’t call these a failure. Not trained to handle such scenario/edge case yet. It is a disengagement, so hopefully you have sent this for future training data.

I've had great luck with v12.5 but then again I don't have any gates. Glad you reported this. My biggest disengagement problem are potholes. Don't want to chance having tire and/or suspension damage. This is especially a problem after the winter thaw when potholes are found more frequently.

FSD did not see the barricade or slow down for it in the slightest. You got lucky that the gate raised, or the Tesla would have crashed right into it .

At what cost is the public providing their lives for future rewards of a personal or corporate profit? When did it become okay to take a life for a product under testing? Shouldn't this remain outside of public harm until proven safe?

Great and very useful information. Thank you

I think it is hard for the system to see such bar It is possible but needs deep training Why not to ask the property to add a sign on the bar It is easier for people as well as the FSD

Camera placement showing all of the steering wheel would be better

Your video title is very misleading.

Can’t believe people here are making excuses about this “It’s not common to have a barricade front of a gate” - like wtf. Don’t run into a barricade, does not matter if there is a gate in front, behind, to the side or nowhere near.

Why don't you repeat the test. Let's see if an intervention is required a 2nd time. The system learns...

I went there and it was beautiful.

Time for a Highland

I love the music

Good test. Still work to be done,but with enough video, should be solvable. Provided the cameras actually see the pole.

Barricade before a gate is definitely a fringe case, give it a while. Interested to see how it handles our just gate in community - seems may be ok.

Excellent edge case! Watch the video in slow motion. @3:19 the speedo says 17mph, @3:21 it has already decelerated to 12mph before the diengagement chime rings. FSD was clearly slowing down before you took over, indicating it saw the gate. I would like to see another attempt

You're still using the old Autopilot on the freeway.

Highway stack is v11 still.

What the heck is a neighborhood gate? Is it to keep the poor people out? This must be the most American thing I've seen today!