Elon: "Feature complete for full self driving this year"

[diplomat33]

I was thinking about how wheel torque driver monitoring isn't feasible for automatic driving on city streets/city NOA given the large turning degree required. How is a driver supposed to keep their hands on the wheel during a 90 deg right hand turn? Is there going to be an alert before any large degree turning actions "Comin' in hot, hands off!"
Consider the case of a driver coming up to a tight 90 deg intersection turn while holding the wheel and not paying attention. Autosteer starts to turn the wheel, driver isn't paying attention and disengages autosteer midturn. Driver refocuses and says oh *** and tries to correct. Highway wheel torque monitoring is pretty straight forward since the wheel degree doesn't go past +/- 45 deg.

I have a feeling that Tesla might be waiting until "Cty NOA" is good enough that the driver won't need to hold the wheel. But I am not sure how that would square with still requiring driver supervision since Tesla uses the torque system. So I am really not sure how Tesla will address this real concern that you bring up.

 

[mt09]

Accidentally disengaging in the inside lane at the start of a double right hand turn would definitely be something you'd want to avoid. On the other hand what's the alternative? Are you really going to be able grab the wheel in time to avoid hitting a car that's only a couple feet away?

The only other option I can think of some really unintuitive time based stuff. Like torque detected within 30 seconds of maneuver if not, abort before performing maneuver. Followed by no torque detected within 5 seconds of maneuver, if torque detected, display warning about upcoming high radial maneuver about to take place. It seems so unintuitive and I couldn't see Tesla releasing it like that. People have a hard enough time with the nags as is. The other option is a much higher torque tolerance during turns to prevent unintended disengagements. They already have dynamic torque disengagement based on some form of confidence and I could see it being applied to this as well. Still doesn't solve accurately monitoring if the driver is paying attention during these higher attention priority moments.

 

[diplomat33]

The only other option I can think of some really unintuitive time based stuff. Like torque detected within 30 seconds of maneuver if not, abort before performing maneuver. Followed by no torque detected within 5 seconds of maneuver, if torque detected, display warning about upcoming high radial maneuver about to take place. It seems so unintuitive and I couldn't see Tesla releasing it like that. People have a hard enough time with the nags as is. The other option is a much higher torque tolerance during turns to prevent unintended disengagements. They already have dynamic torque disengagement based on some form of confidence and I could see it being applied to this as well. Still doesn't solve accurately monitoring if the driver is paying attention during these higher attention priority moments.

Maybe it will be like the current auto lane change with no confirmation in Highway NOA. The car will nag you right before a high turn maneuver, and if you satisfy the torque, the car performs the high turn without requiring hands on wheel during the turn and then resumes nagging after the turn is complete. If you don't satisfy the torque, the car cancels the maneuver and keeps going straight. Not sure if that would work though.

 

[mt09]

I have a feeling that Tesla might be waiting until "Cty NOA" is good enough that the driver won't need to hold the wheel. But I am not sure how that would square with still requiring driver supervision since Tesla uses the torque system. So I am really not sure how Tesla will address this real concern that you bring up.

Yeah I agree. It'll be interesting to see how they handle the transition. If they can get it to a point where the monitoring requirements can be set to much larger intervals then torque monitoring would still be valuable as a crude driver check-in.

 

[mt09]

...If you don't satisfy the torque, the car cancels the maneuver and keeps going straight. Not sure if that would work though.

That's true. I could also see the current failover of putting the hazards on and coming to a complete stop. Especially for 3-way intersections or roundabouts where there's no straight or safe path and require some negotiation to take place.

 

[diplomat33]

That's true. I could also see the current failover of putting the hazards on and coming to a complete stop. Especially for 3-way intersections or roundabouts where there's no straight or safe path and require some negotiation to take place.

Yeah, if the car is already stopped at an intersection, remaining stopped and putting on the hazards if the driver fails the torque test, would be a good idea. Sure it would annoy other drivers but it would be better than the car making a potentially risky turn at an intersection if the driver is not paying attention.

 

[Daniel in SD]

The only other option I can think of some really unintuitive time based stuff. Like torque detected within 30 seconds of maneuver if not, abort before performing maneuver. Followed by no torque detected within 5 seconds of maneuver, if torque detected, display warning about upcoming high radial maneuver about to take place. It seems so unintuitive and I couldn't see Tesla releasing it like that. People have a hard enough time with the nags as is. The other option is a much higher torque tolerance during turns to prevent unintended disengagements. They already have dynamic torque disengagement based on some form of confidence and I could see it being applied to this as well. Still doesn't solve accurately monitoring if the driver is paying attention during these higher attention priority moments.

You could have no disengagement on wheel torque unless you're also pressing the brake or accelerator. To me the issue is that you have to have your hands hovering over the steering wheel ready to take over in tight maneuvers. It's possible city NoA will be so terrifying that nags are not needed

 

[Cirrus MS100D]

Perhaps they’ll just roll out City NOA more slowly in terms of routes and turning radius, similar to “unsupported maneuver coming up” like they do today on highway.

The drive from our house to school has 1 nasty turn, but the others are relatively easy and “shallow.” Maybe City NoA does 90% of them, but hands back control before the tricky ones in initial rollouts?

 

[Buddy]

Given the title of this thread, it will be true no matter when FSD comes out.

Hope people know that people have been speculating this since 2016...

"Any day now!"

 

[CaptMhack]

Unfortunately FSD is too complicated to be done soon.. IMHO it will be years before City FSD is out.. But the good news is that incremental improvements to autopilot are coming in a few month.. Musk is breathing down the software developers neck to get some kind of progress to show.. Which is actually not optimal, I would rather we forget about FSD for a year and let them develop it properly.. Undercooked City FSD will be disastrous in terms of PR - imagine the sophistication of the software to execute a left turn on a 4 lane road during rush hour.. not many human drivers are able to do that now..

 

[powertoold]

Perhaps they’ll just roll out City NOA more slowly in terms of routes and turning radius, similar to “unsupported maneuver coming up” like they do today on highway.

The drive from our house to school has 1 nasty turn, but the others are relatively easy and “shallow.” Maybe City NoA does 90% of them, but hands back control before the tricky ones in initial rollouts?

Just not practical or feasible IMO. The car probably can't predict a turn more than 3 seconds in advance, unless it's dependent on GPS, which may not be accurate.

This whole nagging thing won't work with city NoA.

What I see them doing is allowing the driver to specify a destination, and once the car maps out the route that it intends to take, it will determine whether or not you can enable city NOA. At first, these routes will be very simple, maybe ones that I have already been validated by shadow mode. As Tesla gets more data and is able to validate their approach, more routes will be opened up.

 

[Cheburashka]

I thought it was already feature complete.

They even had a demo day in Palo Alto where they showed it off. There is a 12min video posted by Tesla which shows how it works.

 

[Bladerskb]

The US only has 4 million miles worth of roads, and Waymo is not driving on most of them. How much duplication is there given their current geographic limitations?

Out of that 4 million miles of roads, way less than 5% is actually unique.
Its not about the road topography its about the interaction and the uniqueness of the road.
So out of the 4 million, probably less than 10,000 miles are actually unique.

This is why testing in places like SF, NY, Jerusalem, Phoenix, China, etc... will give you all the data you need because you don't need to test everywhere because the road topography ain't unique. This is why we humans can drive from one state to another.

China

Teslas travel over a billion miles a year with AP active, and most of the fleet can analyze data with it off. Worst case, it takes Tesla 12 months to collect a billion miles of geographicly diverse real world performance data on a new version. (Which also allows for seasonal coverage).

If you ignore all facts and evidence i just laid out, sure.

1. A mile of data would be 1.2 GB - 3.6 GB.
2. Tesla only collect 10 seconds of footage because of ram limits.
3. Data is only uploaded on wifi because of its vast size (10 seconds of footage is 100-300MB).
4. Wifi is only available when you return home. meaning best case of possible scenario is 10 seconds of footage data upload per day.
5. It currently takes a miracle for your car to even get a trigger campaign or have a disengagement trigger a data upload.

With ~400k sales in 2020 and an average yearly mileage of 10,000 , the fleet data collection/ observation will grow by 4 billion miles/ year or a billion miles every 3 months.

• So if you went on a 2,000 miles trip, best case scenario is 10 seconds of footage is uploaded when you return home.
• If you had a yearly commute of 10,000 miles, best case scenario is 4.5 minutes per year from each car.

Teslas travel over a billion miles a year with AP active...
By 2022 there will be a billion miles driven every month (not all AP).
Article on public release of a subset of Waymo's data:
Full Page Reload

Again stop comparing AP with SDC and their data collection.
AP is a glorified lane keeping and adaptive cruise control system.

This is different from what data SDC cars are collecting and how they function.
Its not the same, stop comparing like its the same.

 

[Egga]

Maybe it will be like the current auto lane change with no confirmation in Highway NOA. The car will nag you right before a high turn maneuver, and if you satisfy the torque, the car performs the high turn without requiring hands on wheel during the turn and then resumes nagging after the turn is complete. If you don't satisfy the torque, the car cancels the maneuver and keeps going straight. Not sure if that would work though.

How do Waymo and Uber do it with their test cars and safety drivers? Anyone with any ideas? Do they simply never check driver attention or auto-disengage, and completely trust their safety driver for all needed corrections at all times?

 

[diplomat33]

How do Waymo and Uber do it with their test cars and safety drivers? Anyone with any ideas? Do they simply never check driver attention or auto-disengage, and completely trust their safety driver for all needed corrections at all times?

They probably use a driver facing camera. Pretty much everybody uses that since it allows you to check if the driver is looking at the road without their hands needing to be on the wheel.

 

[DanCar]

They probably use a driver facing camera. Pretty much everybody uses that since it allows you to check if the driver is looking at the road without their hands needing to be on the wheel.

Warning rumor: This is what I heard so likely incorrect or exaggerated. First time Google starting monitoring drivers they were shocked how inattentive some were. They fired 20% at that point.

 

[diplomat33]

Warning rumor: This is what I heard so likely incorrect or exaggerated. First time Google starting monitoring drivers they were shocked how inattentive some were. They fired 20% at that point.

Yes, back in the early days of Google's autonomous driving, they had a L2 system similar to AP and found that drivers do not pay attention. Human drivers get overconfident with a L2 system that seems good enough and start playing on their phones or doing other things.

Here is a video of what Google saw when they "spied" on their drivers:

The CEO of Waymo explained that is why they dropped their work on L2 driver assist completely and focused entirely on full autonomous driving. Basically, they feel that advanced L2+ is fundamentally unsafe because drivers won't pay attention so it is better to do full autonomous driving where the car can do all the driving and you can remove the human element entirely.

 

[banned-66611]

I have a feeling that Tesla might be waiting until "Cty NOA" is good enough that the driver won't need to hold the wheel. But I am not sure how that would square with still requiring driver supervision since Tesla uses the torque system. So I am really not sure how Tesla will address this real concern that you bring up.

For a Level 3 system the car is allowed to notify the driver that they need to take over soon rather than monitoring constantly. Maybe they are hoping to get to Level 3.

 

[mongo]

Illustrative example of a freezer monitoring system:
System 1: Sends temperature reading straight to cloud. Central server checks temperature and sends alert if it exceeds safe value. High data usage.

System 2: Validates temperature readings locally. Sends warning to cloud if temperature exceeds parameters. Low data usage.

No reason to send all the data up.

Wifi is only available when you return home. meaning best case of possible scenario is 10 seconds of footage data upload per day.

Hummmm...
Twitter

It will enable video when parked & connected to WiFi. All Tesla Superchargers will have free WiFi over time.

Waymo currently valued at 30 Billion, may become equipment provider:
https://www.forbes.com/sites/samabuelsamid/2020/03/06/waymos-30b-valuation-shows-the-new-reality-of-automated-driving-is-sinking-in

 

[SOULPEDL]

The CEO of Waymo explained that is why they dropped their work on L2 driver assist completely and focused entirely on full autonomous driving. Basically, they feel that advanced L2+ is fundamentally unsafe because drivers won't pay attention so it is better to do full autonomous driving where the car can do all the driving and you can remove the human element entirely.

Which could also be a convenient way of hiding the fact that they can't do this blended mode with humans in a timely fashion (especially w/o data volume). So it could be a punt in hopes that the other team (Tesla) fumbles before Waymo removes their steering wheel first? Or at least that could be their pitch to investors. "We're taking the safer route to full autonomy" I can imagine being said, which is clearly the endpoint in everyone's mind, hoping to leapfrog Tesla with this shortcut, and avoid lawsuits along the way. This makes the most sense in my head.