FSD Beta Videos (and questions for FSD Beta drivers)

[diplomat33]

40K miles per accident is a number google / Waymo reported many years ago.

Can you be more specific? What type of accident does the 40k miles stat include?

I am guessing since the number is so low that it is counting all incidents, even minor ones. But it should be noted that counting minor accidents is difficult since a lot them are not reported to law enforcements.

 

[rijc99]

I thought of that too in previous videos but did not mention it. I think a major problem (which has been brought up before) is Tesla's nav does not allow waypoints. This is a common complaint that has been brought up endlessly. If the nav allowed waypoints, all FSD has to do is set one up a waypoint that avoids that left turn, and it should be fine.

Of course the other question of adding a place to avoid in the navigation, but I don't think most navigation systems allow that. However, I think behind the scenes it should be possible to shoehorn the traffic module to do that (just set the intersection as heavily congested and nav should avoid it).

It was weird when FSD got into the endless right turn loop because every other navigation systems I’ve seen would’ve rerouted the path to include an u turn at the light instead of doing a circle.

 

[Chazman92]

#FSDBeta 9.1 - 2021.4.18.13 - Forward Facing Unprotected Left Turns with Drone View - Much Better

Third video of the day on v9.1 check it out.

 

[powertoold]

#FSDBeta 9.1 - 2021.4.18.13 - Forward Facing Unprotected Left Turns with Drone View - Much Better

Third video of the day on v9.1 check it out.

Still unacceptable, but it does seem like it sees all the cars. The creeping and commit behavior is still dangerous though.

edit: it doesn’t actually see the cars from afar based on what’s shown on the visualization, and that’s why it starts creeping or committing in a dangerous manner

 

[diplomat33]

Still unacceptable, but it does seem like it sees all the cars. The creeping and commit behavior is still dangerous though.

I wonder if that is one the "fundamentals" that Elon says they will need to work on before they can release V10? If not, what do you think the "fundamentals" are that Tesla still needs to work on?

https://twitter.com/x/status/1421564618258931715

 

[powertoold]

I wonder if that is one the "fundamentals" that Elon says they will need to work on before they can release V10? If not, what do you think the "fundamentals" are that Tesla still needs to work on?

https://twitter.com/x/status/1421564618258931715

Elon isn't in the trenches developing fsd, so he's just talking to us at a high level. But Tesla needs to improve perception even further... I think the improvements in 9.1 are attributable to improved / increased perceptual confidence (as shown by the bolder and sharper lines in the visualization vs 9.0). The production pure vision fleet is still limited to 80mph, so there are still deficiencies in pure vision perception.

Tesla is still far away from wide release IMO, but the progress in last three weeks is promising, as others have mentioned.

 

[diplomat33]

Elon isn't in the trenches developing fsd, so he's just talking to us at a high level. But Tesla needs to improve perception even further... I think the improvements in 9.1 are attributable to improved / increased perceptual confidence (as shown by the bolder and sharper lines in the visualization vs 9.0).

Tesla is still far away from wide release IMO, but the progress in last three weeks is promising, as others have mentioned.

Thanks.

My early impressions is that Tesla is making progress with vision and FSD beta is getting to the point where it can do easy routes pretty well. I define easy routes as any drive that does not have a lot of road users. I agree Tesla needs to improve perception further. But the real challenge of FSD is prediction, planning and driving policy. That is what is needed to actually do complex driving reliably. As I see it, FSD on empty roads is trivial. What makes FSD a difficult problem to solve is the presence of other road users because they are independent actors. They can be hard to predict. And good and safe driving requires that the FSD system juggle navigating a route, obeying traffic rules, driving smoothly, avoiding collisions, and respecting other road users, all at the same time. That is why even if you solve perception, you still have a lot of work to do in order to solve the last 10% of FSD because the real challenge is training the car to make smart decisions about how to react to other road users in a huge number of driving cases and environments.

 

[Daniel in SD]

It's always odd to me when people say the display inspires confidence. To me it's the exact opposite because the visualized distance is so short. In this example it doesn't show a full size truck that's 200ft away. Not only that it appears not to see it all and attempts to cross in front of it.

 

[powertoold]

As I see it, FSD on empty roads is trivial. What makes FSD a difficult problem to solve is the presence of other road users because they are independent actors. They can be hard to predict. And good and safe driving requires that the FSD system juggle navigating a route, obeying traffic rules, driving smoothly, avoiding collisions, and respecting other road users, all at the same time. That is why even if you solve perception, you still have a lot of work to do in order to solve the last 10% of FSD because the real challenge is training the car to make smart decisions about how to react to other road users in a huge number of driving cases and environments.

I can see why you think that, but with Tesla's approach, FSD on "empty" roads is not trivial at all. When you say "empty," you're referring to other dynamic objects, but it's not at all trivial (Tesla's approach) to understand what objects are dynamic and/or relevant. For example, there are many cars parked on the road, all sorts of lights and objects, things that look like traffic lights, etc. V9 needs to understand and take into consideration all of these things as it's driving. Other FSD developers have likely concluded that it's "impossible" for them to deal with all these objects and decided to use HD-maps to label / hand verify which objects / lanes are relevant.

Plus, policy and planning are as relevant on empty roads as they are with other actors, as many of V9's issues stem from odd lane management or choices.

There are many aspects of V9 that we've simply taken for granted but are magical if you consider what's going on. For example, here, where V9 sees the parked car, correctly understands that the oncoming car isn't a collision threat, gives it room to pass, and then passes itself. Chuck doesn't even mention it, because it's "routine" now. It's also something I've yet to see Waymo do in the videos, as it takes a high level understanding that an oncoming car can be in my lane without being a threat:

 

[mhan00]

#FSDBeta 9.1 - 2021.4.18.13 - Forward Facing Unprotected Left Turns with Drone View - Much Better

Third video of the day on v9.1 check it out.

I have no words for the horrific decision by the car at 1:45. Of all the decisions the car could have made, it made the very worst.

 

[n.one.one]

This is promising.

 

[diplomat33]

I can see why you think that, but with Tesla's approach, FSD on "empty" roads is not trivial at all. When you say "empty," you're referring to other dynamic objects, but it's not at all trivial (Tesla's approach) to understand what objects are dynamic and/or relevant. For example, there are many cars parked on the road, all sorts of lights and objects, things that look like traffic lights, etc. V9 needs to understand and take into consideration all of these things as it's driving. Other FSD developers have likely concluded that it's "impossible" for them to deal with all these objects and decided to use HD-maps to label / hand verify which objects / lanes are relevant.

You are correct that it is probably not as trivial for Tesla but that is because of Tesla's vision-only approach. With vision-only, you have to do the entire perception, scene understanding, prediction, planning. So vision-only has to solve all FSD problems by itself. And if your vision is missing anything, it can problems which make FSD less safe. To be clear, I am not saying that vision-only FSD is impossible. But it is a lot harder IMO.

Yes, other AV companies use HD maps and other sensors like lidar and radar precisely for this reason. I don't think they consider it "impossible" without them. It's just that using HD maps and other sensors will make the job a whole lot easier and much more reliable. When it comes to "solving FSD" why wouldn't you make the job easier and more reliable? After all, the goal of FSD is to do it safely, consistently and reliably every time. IMO, you should use whatever tool can help you get the job done. Why would you not use every tool that can help you solve FSD? That is why I am a fan of using HD maps, lidar and radar in addition to cameras.

And we see the results: Waymo, Cruise, Zoox etc are very reliable at those "basic" tasks like picking lanes and not hitting other objects things that Tesla still has issues with the vision-only approach.

Plus, policy and planning are as relevant on empty roads as they are with other actors, as many of V9's issues stem from odd lane management or choices.

Policy and planning will be much simpler on empty roads than with other actors. On empty roads, you basically only need to stay in the lane, avoid static objects, follow basic traffic rules like stop lights and road signs, and make turns to follow the right lane path. But you don't have to worry about hitting other road users. But with other roads users, you have to add on top of that, predicting their paths, avoiding hitting other road users when lane changing or that might cut in front of you, deciding when to lane change to respect other road users, yielding to other road users etc...

 

[diplomat33]

I have no words for the horrific decision by the car at 1:45. Of all the decisions the car could have made, it made the very worst.

That's disappointing.

 

[diplomat33]

There are many aspects of V9 that we've simply taken for granted but are magical if you consider what's going on. For example, here, where V9 sees the parked car, correctly understands that the oncoming car isn't a collision threat, gives it room to pass, and then passes itself. Chuck doesn't even mention it, because it's "routine" now. It's also something I've yet to see Waymo do in the videos, as it takes a high level understanding that an oncoming car can be in my lane without being a threat:

If you are referring to the situation at 3:20 in the video with letting the car pass, Waymo can handle that.

Here is a video from 3 years ago, where we see Waymo handle a similar situation where it let's an oncoming car pass first in a parking lot with parked cars and pedestrians. At the 2:56 mark:

In fact, I would recommend you watch the entire video as it is informative of how the Waymo Driver works, the kind of perception and planning and driving policy that it has. But keep in mind, the video is 3 years old. So the Waymo Driver today is probably way better than what you see in the video. But it will give you an idea of what the Waymo Driver could do 3 years ago.

Also, remember that there are a lot of cases that Waymo can probably handle that are not captured on video because most of Waymo's autonomous driving is outside of Chandler that the public does not get to see since it is still in the testing phase. The public is only seeing a very small fraction of Waymo's autonomous driving.

 

[powertoold]

If you are referring to the situation at 3:20 in the video with letting the car pass, Waymo can handle that.

Here is a video from 3 years ago, where we see Waymo handle a similar situation where it let's an oncoming car pass first in a parking lot with parked cars and pedestrians. At the 2:56 mark

It's a total Waymo marketing video. I value actual user videos way'mo. When I said video, I was more referring to JJRick's / others ride videos. If you know me, I tend to take marketing videos with a grain of salt.

Also, it's not the same situation, the oncoming car wasn't in the Waymo's lane. The speeds are slower, etc.

 

[rijc99]

Waymo has the benefit of alot more data to make decisions. Hopefully Tesla can work around the limitations of vision only.

 

[diplomat33]

It's a total Waymo marketing video. I value actual user videos way'mo. When I said video, I was more referring to JJRick's / others ride videos.

Also, it's not the same situation, the oncoming car wasn't in the Waymo's lane.

I know what you mean. But it is still a video of actual autonomous driving. They just overlayed some extra graphics to explain what the car is thinking. The autonomous driving in the video is real.

And IMO, the Waymo case was more difficult than the Tesla case. The Waymo had to maneuver in a narrow space in a parking lot with pedestrians as well as letting the car pass. The Tesla was on a residential street with just a few parked cars.

And I know you want videos from real users. But right now, the only Waymo videos from real users will be from Chandler which won't tell us how Waymo handles driving in other areas. We don't have videos from real users from Waymo driving in other areas since Waymo has not opened it up to the public yet. So demanding videos from real users does not really help until Waymo opens up their autonomous driving in other areas to the public.

 

[nativewolf]

The lack of real user videos is emblematic of the slow rate of progress and the general failure of the program- @powertold is absolutely correct to be skeptical of any and all cherry picked marketing videos. Excusing everything Waymo is not doing but criticizing everything that Tesla is doing is not really pushing the conversation forward- instead ask why won't Waymo let more video be released in more towns. What Tesla did is make proper decisions re a potentially dangerous situation with only camera based data at speed from a real person in a real car somewhere in NA. In driving it is speed and uncertainty that kills, slow things down enough in a small box and all becomes simple.

Tesla is all in a platform that could actually be rolled out tomorrow if Tesla wished to do so. The hardware exists, the platform exists, the market is demanding product, pricing is viable, and they have almost a million testers. When Tesla says ok, this is good enough then anyone with $10k can have some degree of a chauffeur- and one that will get better with age. Waymo is, as of this moment, not something that can actually be rolled out. It is a giant beta, neither hardware, not software, nor operating ecosystem, as they exist today, will ever make it to the general public. It is just a giant Google experiment going the way of most google experiments- neat and push boundaries but no urgency or actual idea of how to deliver a viable product to market. Why do you think all the leaders have bailed? Googles wins (after initial launch) have almost all been related to buying smart and well run little companies and scaling them up. Nearly everything developed inside the borg, not great. I've actually never even taken Waymo seriously just because it is a google project with no viable path to market. If google did not understand how to market something than why go to the trouble in the first place. For google engineers it is a fun problem to solve ...for the CFO it is a giant cash sucking pit. Maybe Tesla solves FSD enough to release and Waymo is sold to Toyota or VW who are both desperate to not get crushed and have the ability to raise $30 billion or more and google gets $. Who knows. Waymo had to take in another $2.5 billion just last month because they are such a cash sucking parasite. As it is today I see no possibility that Waymo ever succeeds- google has no idea how to productize it and it is losing talent; if it is sold it will get starved of talent and die a different death.

Waymo and the vast expenditure and terrible rate of progress would leave me without hope. I do have hope that Tesla will actually solve the conundrum to a degree sufficient to allow me to have my truck (ct) take me to work while I look at trees out the window.

 

[rijc99]

I think both are making good progress. Neither are complete solutions at this moment. The more companies that succeed in this self driving car venture the better it is for consumers. Last thing we want would be to only have one company run the whole thing.

 

[AlanSubie4Life]

I do have hope that Tesla will actually solve the conundrum to a degree sufficient to allow me to have my truck (ct) take me to work while I look at trees out the window.

You’ll be living the dream! Looking at trees, while trying to avoid looking at videos of FSD 9.1.