What statement was inaccurate? I didn’t comment on whether FSD or Waymo’s system was “superior.” Just that Tesla’s system (which is Level 2) requires a human driver to be driving at all times, while Waymo’s does not require a human driver to be driving in order to operate safely. Waymo still has interventions for the car to get to its destination, because a Level 4 system doesn’t necessarily guarantee that the car can drive to all places - just that it can always be operating safely without a human while it is driving. FSD isn’t that kind of system yet. And Musk isn’t going to announce on 10/10 that it’s reached that point.
3 Likes
Point being there is almost no basis for comparison. Waymo is in a very restricted, closely defined environment. Tesla is in the wild. Waymo can pull to the side, ask for help, and not count that as an intervention. Tesla FSD’s only choice is to say “driver take over now” and I’ll bet that counts as an intervention. Waymo has no one in the car to decide to go faster, take a different route, or whatever, while Tesla always does. They are really just not directly comparable.
2 Likes
Oh, I agree they’re not comparable. Waymo has reached the point where the car never says, “driver take over now” (which is what Level 4 is), but for a limited geofenced area. Tesla’s Level 2 system can be operated in nearly every all environments, but has not reached the point where there does not have to be a human driving at all times.
If Tesla wants to have a robotaxi or full self-driving, however, they need to get at least to where Waymo is - the point where the car can operate safely without a human driver, at least while it’s driving. There’s nothing that suggests they’re anywhere close to being able to do that - they haven’t sought permission to operate a Level 4 system, they don’t publish their data, and even the most favorable crowd-sourced data is orders of magnitude worse than where they need to be to remove the driver from the driver’s seat.
Maybe Tesla will provide some information on FSD performance on 10/10 that sheds light on this. But right now, there’s no reason to think that they’re going to be in a position to operate a driverless system any time soon.
5 Likes
The point being made is that Waymo is not nearly as advanced as you think they are. The references to the remote driver are quite frequent at this point so they are not driving huge distances without human input.
These assertions are flawed. Cruise reached that point many months ago and then GM stopped the program.
We came to know later that there were too many remote interventions.
Waymo is better than Cruise but in my opinion, it is a fundamentally flawed approach. There is a reason why after Google deployed $Billions has not been able to scale this after decades of work.
I don’t think anyone is going to crack full self driving and robotic taxis. Nobody. October 10 is going to be a disappointment. And many companies will have wasted a ton of money and resources on this.
1 Like
Maybe Waymo has a long way to go as well - but the point still stands. In order to operate a driverless system, the cars when they are driving have to be able to operate with a very low frequency of incidents that require immediate human intervention. In order to be economical they may also require a low frequency of non-critical incidents that do not require immediate human intervention, but require intervention nonetheless. But they absolutely have to meet the first standard as well.
I’m not claiming that Waymo has advanced enough to operate an economically viable, scalable robotaxi service. But they have reached the point where their cars can safely operate for a hundred thousand trips a week without having to be driven by humans. They are driving huge distances without human drivers. They still require human input in non-driving situations - perhaps an enormous amount of input - but when the cars are driving, they’re driving without human intervention. Tesla, OTOH, has not reached that point.
Note - that doesn’t mean FSD isn’t an amazing ADAS system, because Tesla is trying to do something harder than Waymo is trying to do. They’re aiming for Level 5 (or a vastly broader version of Level 4). But there’s no reason to think that they’re close to being able to achieve that goal, because all of the (admittedly uncertain) measures of their critical intervention rates are orders of magnitude worse than they need to be to operate without a driver.
The Waymo human drivers are not in the car but humans are controlling it remotely. If your point is that Tesla does not even have this, then it is a valid.
The following is not a Waymo control room, but it is something like this.
**strong
Perhaps - but that doesn’t mean that Tesla is close to solving it, either. Waymo is an illustration of the type of performance you need to be able to deliver in order to remove the human from behind the steering wheel. They might have pursued a flawed approach to get there, but you can’t operate a Level 4 system unless you get your critical intervention rate down to those levels. The average American driver only gets in a crash once per 670K miles, so a Level 4 system really has to get pretty close to that level of critical interventions (ie. an intervention to avoid a crash, not an intervention for some ‘discretionary’ reason like preferred driving style or taking an alternate route).
Waymo has reached that point. They’ve reached that point for a limited scope of service, and it still requires non-trivial human involvement outside of critical situations - but they have achieved the baseline level of performance of the car being able to operate without a human driver. Tesla has not.
2 Likes
No, they’re not controlling it remotely in that manner. Waymo operators do not have the ability to drive the car in place of the AV. They have the ability to tell the AV system what to do, but they’re not taking over remote control of all driving functions the way a human driver in a vehicle would. The car is always driving, but they are “asking questions” of the remote humans.
Here’s Waymo’s description.
2 Likes
Agree. Tesla is on its way but not there yet. However the approach is rock solid and the rate of progress is very high.
There are some more breakthroughs needed
- FSD does not have a concept of memory. It drives on the road like its driving for the first time and does small thing differently
- FSD does not have the concept of “car queued” vs “car stopped”
- On highways while changing lanes, it does not identify a gap (like a human) between cars, smartly speed or slow down to that gap and then merge.
I have used FSD for 8 years now. It can do hundred of things that it could not do earlier. I used to wonder how they would do overcome those problems but they solved most of them.
Is it very high? Is the rate of progress high enough for actual autonomy?
Again, Tesla doesn’t release official data on FSD performance. Not a knock on them - that’s an appropriate business move. The crowdsourced data shows a system that’s several orders of magnitude away from being able to operate without a human driver. They’ve had material progress in the immediate past, but they’ve already reached the point where they’re getting fewer and fewer useful scenarios from the firehose of data they’re processing.
So, are they in a position to improve their performance 1000x from where they are? Within shorter than a few years? No indication that’s on the horizon - or more likely to get a result than Waymo expanding their geofence further and further…
1 Like
Last few OTA installs v11 → v12.3 → v12.4 → v12.5 have been within a span of weeks. I am on HW3. Newer cars are on HW4 (better cameras, better inference chips). HW5 coming in 2025.
More cars, more data, more training, faster OTA updates. Moreover, FSD going global next year (Europe, Canada, China).
So yes, rate of progress is fast.
A key issue here is the definition of “it”. Waymo is clearly targeted at robotaxi behavior and is willing to accept some fairly severe restrictions in geofencing and speed in order to achieve that goal. Tesla is also targeting robotaxi behavior, but is not willing to accept those restrictions, but is happy to deliver cars which can successfully drive by themselves in increasingly many environments. Given the complexity of the differences between the two strategies, I doubt that we are in a position to really compare progress toward goal, especially since we now know that there is much more human involvement than was apparent from merely having no human driver in the car.
Err, the accident rate for FSD is already much lower for FSD than an average driver. Not close to the level, but well beyond it.
In a 2022 study … eons ago for FSD
“Vehicles equipped with FSD Beta crash only 0.31 times per million miles on non-highway roads, compared to the national average of 1.53 crashes per million miles for human drivers” .
1 Like
It’s a nonsense statistic. Most FSD takes place on limited access roads, where the accident rate is already well below average. Moreover, it’s self-selected; in hazardous conditions drivers don’t choose to use it. Moreover it doesn’t include cases where the human takes over control in a dangerous situation. It’s a simple recitation of a statistic, without context and without meaning.
But do carry on.
I thought the idea of “robotaxi” was, you know, “taxi”. There are precious few taxi rides which go from one city to another, so for probably 99.9%, they are “geofenced” in reality, if not in regulation.
3 Likes
What is a “limited access road” exactly? And why do you assert this? I’ve used FSD everywhere, local roads, highways, mixed, airport runs, long trips, etc.
Only in some hazardous conditions. A few months ago, on a trip back south to home, during a very heavy rainstorm, I chose to use FSD because apparently it was seeing the lane lines a lot better than I could see them. The other choice was to pull over and wait for the storm to pass and that would have been somewhat dangerous due to possibility of being hit from behind on the shoulder by a driver thinking my lights are where the lane is.
The way I understand it is that when they say “robotaxi”, they mean TAAS in general. And in the extreme, they mostly mean “not owning a vehicle anymore”, so that includes ALL vehicle usage, taxi rides 2-15 miles, airport runs (10-75 miles mixed city/highway), AND also longer rides from city to city on highways.
I don’t particularly like the word “robotaxi”. Of course, I also don’t like the words “autopilot” or “FSD” either. 
I would prefer words with much better clarity than those.
2 Likes
That’s a well defined term. No stoplights, entrance and exit ramps. Like the Interstates, toll roads, etc. Inherently safer driving because traffic mostly flows, rather than stop, start, and cars entering, running red lights, etc.
Because Tesla once said it’s the sum total of all miles driven with FSD divided by accidents. Other than that they are terribly opaque about their methodology. The NHTSA is not, however, and doesn’t have a dog in the fight:
When FSD was introduced in 2014 it was limited to “ramp on, drive, ramp off” according to the company. I’m aware that has changed, but including that kind of driving
“Tesla is having more severe — and fatal — crashes than people in a normal data set,” she said in response to the figures analyzed by The Post. One likely cause, she said, is the expanded rollout over the past year and a half of Full Self-Driving, which brings driver-assistance to city and residential streets. “The fact that … anybody and everybody can have it. … Is it reasonable to expect that might be leading to increased accident rates? Sure, absolutely.”Cummings said the number of fatalities compared with overall crashes was also a concern.
It is unclear whether the data captures every crash involving Tesla’s driver-assistance systems. NHTSA’s data includes some incidents in which it is “unknown” whether Autopilot or Full Self-Driving was in use. Those include three fatalities, including one last year.
https://www.washingtonpost.com/technology/2023/06/10/tesla-autopilot-crashes-elon-musk/
I didn’t mean it as a sweeping ‘nobody ever uses it” but it’s clear that a lot of people take over the driving in unusual circumstances, and that makes Musks’s data useless through self-selection-bias.
I suspect it will be a very long time before that vision comes true. The sheer size of fleet you need to effect that means a huge investment in metal and batteries to equalize the fleet. The car & truck rental companies didn’t make that leap for 20 years. It’ll be faster now, but there are a lot of hurdles before that becomes reality.
1 Like
The US Interstate Highway System is a “limited access highway”. There are relatively few on ramp and exit ramps, and almost all of the highway alignment is fenced off.
intercst
Okay. But that ISN’T how FSD is used at all. The way you use FSD is by entering a destination into the navigation system, then the system routes to that destination, then you activate FSD, and then it drives you there. That includes all kinds of roads, stoplights, stop signs, turns, entrance/exit ramps, etc.
Your limited access comment would apply to autopilot. You don’t enter a highway and then activate FSD, that’s what they call autopilot (basically a fancy way to describe lane keeping, speed keeping, and distance keeping, aka a glorified cruise control).
2 Likes
“It” meaning an autonomous system that can operate without a human driver behind the wheel. That’s what Tesla needs to achieve in order to have a robotaxi. You’re right that they’re pursuing other goals as well, which is almost certainly going to make it take longer for them to get to that milestone. But they still have to hit that milestone. They can’t have a robotaxi without doing that.
Yes, but FSD is not a Level 4 system. It achieves that accident rate with a driver. Because FSD has such a high rate of critical interventions, if there wasn’t a human driving the car (just using FSD as an assist), the accident rate would be catastrophically high. A driverless system has to be safe enough that it can have that super low accident rate (>500K miles between accidents) without human intervention in critical situations. Tesla is far from there with FSD. By the limited accounts we have, it’s several orders of magnitude away from there.
That’s the point. Waymo has developed a Level 4 system that has a driverless accident rate that’s low enough to function safely. It’s geofenced and limited and requires some degree of human intervention outside of driving situations, but it is capable of operating safely without a human driver. Tesla’s FSD appears to be very far away from that milestone, and cannot operate a robotaxi system until they reach the point where it can safely operate without a human driver.
2 Likes