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

[willow_hiller]

This will probably be a "smart-summon" situation again. Where in theory the feature sounds really cool, but in practice it's a little lame.

But hopefully these are just starting points; and will be valuable for Tesla to learn whether their sensor suite is sufficient or whether they need to iterate. For all the valid criticisms of Elon Musk, I don't think he's one to rest on his laurels.

 

[EVNow]

The manual has a long list of disclaimers of where traffic light response won't work:

"Stopping at traffic lights and stop signs is not designed to stop at:
- Railroad crossings
- Keep out zones
- Toll booths
- Crosswalk systems
- Yield signs or temporary traffic lights and stop signs (such as at construction zones)
- Miscellaneous traffic u-turn lights, bicycle and pedestrian crossing lights, lane availability lights, etc...

Interesting list. So, they have all these scenarios they are yet to work on.

In addition, stopping at traffic lights and stop signs is particularly unlikely to operate as intended, can disengage, or may not operate when one or more of the following conditions are present:
- Visibility is poor (heavy rain, snow, etc) or weather conditions are interferring with camera or sensor operation.
- Bright light (such as direct sunlight) is interfering with the view of camera(s)
- A camera is obstructed, covered, damaged or not properly calibrated.
- Driving on a hill or on a road that has a sharp curves on which the cameras are unable to see upcoming traffic lights or stop signs.
- A traffic light, stop sign, or road marking is obstructed (for example, a tree, a large vehicle etc)
- Model 3/Y is being driven very close to a vehicle in front of it, which is blocking the view of camera

None of these are very surprising - but gives us a list of conditions where the error rate is high I their tests, I guess.

When Musk said you can't sleep in the back of the car, he wasn't kidding

 

[diplomat33]

This will probably be a "smart-summon" situation again. Where in theory the feature sounds really cool, but in practice it's a little lame.

But hopefully these are just starting points; and will be valuable for Tesla to learn whether their sensor suite is sufficient or whether they need to iterate. For all the valid criticisms of Elon Musk, I don't think he's one to rest on his laurels.

These limitations mentioned in the manual confirm my concerns about the hardware not being good enough:
- Visibility is poor (heavy rain, snow, etc) or weather conditions are interfering with camera or sensor operation.
- Bright light (such as direct sunlight) is interfering with the view of camera(s)
- A camera is obstructed, covered, damaged or not properly calibrated.
- Driving on a hill or on a road that has a sharp curves on which the cameras are unable to see upcoming traffic lights or stop signs.
- A traffic light, stop sign, or road marking is obstructed (for example, a tree, a large vehicle etc)
- Model 3/Y is being driven very close to a vehicle in front of it, which is blocking the view of camera

If traffic light response can't work reliably in these situations, how can it ever do safe L5 without driver supervision? So yeah, I think that Tesla will need to upgrade the sensor suite if they want to do true safe L5. But in the mean time, we will get good L2 driver assist features.

 

[willow_hiller]

If traffic light response can't work reliably in these situations, how can it ever do safe L5 without driver supervision? So yeah, I think that Tesla will need to upgrade the sensor suite if they want to do true safe L5. But in the mean time, we will get good L2 driver assist features.

Let's assume that L5 is impossible with HW3. At that point their options are:

1. Rip the bandaid off. Stop development of AP with the current sensor suite, potentially pay a lot of refunds or develop a costly retrofit that will need to be furnished to FSD purchasers for free, and start from square one with the new sensors.

2. Try and fulfill the promises made with the current hardware, release most of the features as L2 and hope that makes enough people happy to avoid refunds and retrofits. And then after all that start from square one with the new sensors.

I'm honestly not sure which is worse.

 

[diplomat33]

Let's assume that L5 is impossible with HW3. At that point their options are:

1. Rip the bandaid off. Stop development of AP with the current sensor suite, potentially pay a lot of refunds or develop a costly retrofit that will need to be furnished to FSD purchasers for free, and start from square one with the new sensors.

2. Try and fulfill the promises made with the current hardware, release most of the features as L2 and hope that makes enough people happy to avoid refunds and retrofits. And then after all that start from square one with the new sensors.

I'm honestly not sure which is worse.

I think #1 is worse. And Elon is so adamant that the cars are "FSD capable", there is no way that he will do #1. It would be a complete admission of failure. Elon will never do that. Plus, it would cost a ton of money, not just in refunds but in loss revenue and time from all the AP development going to waste.

#2 is the better option IMO. It would give some "FSD" to current owners. Tesla could still claim partial victory (even if they don't actually achieve L5). And then introduce a new sensor suite that will get a bunch of new sales because you know a bunch of owners will want to get the new hardware. Heck, if the new sensor suite was capable of L5 for real, I'd trade in my Model 3 and upgrade in a heart beat.

 

[anthonyj]

That's it. FSD is a huge waste of money. $7k is better spent on $1500 calls that expire in a week. At least there is a .05% chance of getting something out of those

 

[linux-works]

$7k is better spent on $1500 calls that expire in a week.

who are *you* buying from? wow, around here, the phone cards are good for at least a year.

 

[Daniel in SD]

The slowing down at green lights seems horrible. I wonder why they don't have it beep or vibrate the steering wheel to ask for confirmation? Then allow a confirm with a tap of accelerator that would not cause the car to accelerate.
Overall the feature seems well thought out even though I would never use it (I don't use NoA either though). Hopefully it will be safe. There's still the issue of missing unmapped intersections. I wonder if they get the map data from the fleet?
One interesting note is that it won't let you accelerate into an intersection if the light is red. I wonder how it deals with right turns on red?

 

[Dutchie]

I wonder if this software update doesn't nessecitate driver visualization for the driver to know what FSD actually is seeing. How is Tesla going to do it with HW2 upgrade to HW3 without upgrading MCU1 to MCU2?

 

[diplomat33]

I'm going to make a prediction that "City NOA" will be very similar to this "Traffic Light Response" feature. Based on GPS and some vision, the car will auto lane change to follow nav route (like current NOA) and make turns at intersections but the driver will need to confirm with the stalk before each turn.

 

[mongo]

Well, we got a description from Tesla.

Honestly, this sounds outright dangerous.

Courtesy of @greentheonly Twitter account.

Stopping by default before evey intersection is arguably the safest option. (He argued)

...
So Tesla is releasing a feature that requires active driver supervision, but is not providing proper driver monitoring (no driver facing camera) and saying "oops the feature may fail but you need to pay attention". That is terrible! Tesla is being incredibly irresponsible with this feature.

How is requireing direct input from the driver on each intersection not driver monitoring??
Every other car doesn't require anything before entering a crossing, Tesla requires a conscious state change of pressings the pedal or stalk.

So it’ll stop for all traffic control devices. Except the ones where it won’t. And you have to be ready to press the accelerator, as well as the brake, at all times.

10 year old me would be so disappointed by what the future looks like. Me from a couple years ago is similarly disappointed by what his $8000 will theoretically (AP2.0 representing!) get him.

Yes, being in control/ monitoring operation requires you to be ready to accelerate, brake, or steer at all times. That is what the manual tells you over and over.
It's not FSD at this point.

I really wonder about pressing the accel, to say 'ok'.

just not sure this is fully thought out. I'm thinking this is the best compromise they could come up with, but I have doubts if its going to be the way it is, going forward.

confirm to CONTINUE driving thru green? wow. that's really legal CYA oriented, to me; and not very driver-friendly.

It's meant to be safe (everyone else friendly), not driver friendly. You can also use the stalk to confirm.

its not the ideal thing, but its what *can* be done with what is available today. and you hope that computers A and B agree with 'lots of nines', so that the pullover and stop routine almost never gets called in real life driving. still, you have to plan for it via design and code and testing.

In redundant mode, chips A and B run the exact sane code with the exact same inputs. If the outputs don't match 100%, it's a fault.

But hopefully these are just starting points; and will be valuable for Tesla to learn whether their sensor suite is sufficient or whether they need to iterate. For all the valid criticisms of Elon Musk, I don't think he's one to rest on his laurels.

Yes, it's a wide roll out test to gather more edge cases with driver provided labeling.

These limitations mentioned in the manual confirm my concerns about the hardware not being good enough:
- Visibility is poor (heavy rain, snow, etc) or weather conditions are interfering with camera or sensor operation.
- Bright light (such as direct sunlight) is interfering with the view of camera(s)
- A camera is obstructed, covered, damaged or not properly calibrated.
- Driving on a hill or on a road that has a sharp curves on which the cameras are unable to see upcoming traffic lights or stop signs.
- A traffic light, stop sign, or road marking is obstructed (for example, a tree, a large vehicle etc)
- Model 3/Y is being driven very close to a vehicle in front of it, which is blocking the view of camera

If traffic light response can't work reliably in these situations, how can it ever do safe L5 without driver supervision? So yeah, I think that Tesla will need to upgrade the sensor suite if they want to do true safe L5. But in the mean time, we will get good L2 driver assist features.

They use the map (memory) to know there is an intersection ahead, just like anyone else would.

These warnings are due not having the data to prove reliability. It does not mean it always fails in those conditions. Besides, name a sensor Tesla isn't using that can see around corners or through obstacles.

The slowing down at green lights seems horrible. I wonder why they don't have it beep or vibrate the steering wheel to ask for confirmation? Then allow a confirm with a tap of accelerator that would not cause the car to accelerate.
Overall the feature seems well thought out even though I would never use it (I don't use NoA either though). Hopefully it will be safe. There's still the issue of missing unmapped intersections. I wonder if they get the map data from the fleet?
One interesting note is that it won't let you accelerate into an intersection if the light is red. I wonder how it deals with right turns on red?

The UI pops up a red bar to alert you that it will brake in the future (before it starts slowing). Then you tap the accelerator or down on the stalk to confirm.

Same deal at NoA lane changes initally or train driver attention checks.

 

[ZeApelido]

This implementation makes sense. Everyone knows there was no way they were going to have an implementation that was not going to require user oversight. Otherwise plowing through red lights was going to happen. Even when noting its drivers responsibility, news would not put Tesla in a good light.

This allows them to start collecting massive amounts of labels. Every update can raise the probability threshold a bit as the model improves, allowing sensitivity to stay high but reducing false negatives. Over time, the cars will not slow down in the clearest, most obvious green light conditions.

The real question will be, how far can they take this in say a year or so?

 

[diplomat33]

How is requiring direct input from the driver on each intersection not driver monitoring??
Every other car doesn't require anything before entering a crossing, Tesla requires a conscious state change of pressings the pedal or stalk.

Yes, driver input is a form of driver monitoring. I was thinking of how Tesla does not use a driver facing camera which could allow Tesla to monitor the driver without needing direct driver input.

They use the map (memory) to know there is an intersection ahead, just like anyone else would.

I don't fault Tesla for using map data. But it is funny how Tesla fans say that HD maps are a crutch. Yet, Tesla uses map data.

These warnings are due not having the data to prove reliability. It does not mean it always fails in those conditions. Besides, name a sensor Tesla isn't using that can see around corners or through obstacles.

Well, Waymo has perimeter sensors that help see around corners. And the whole point of HD maps is to tell the car where stuff is when the sensors might not be able to see it. In fact, extra sensors and/or HD maps would solve all these limitations:

- Visibility is poor (heavy rain, snow, etc) or weather conditions are interfering with camera or sensor operation.
- Bright light (such as direct sunlight) is interfering with the view of camera(s)
- A camera is obstructed, covered, damaged or not properly calibrated.
- Driving on a hill or on a road that has a sharp curves on which the cameras are unable to see upcoming traffic lights or stop signs.
- A traffic light, stop sign, or road marking is obstructed (for example, a tree, a large vehicle etc)
- Model 3/Y is being driven very close to a vehicle in front of it, which is blocking the view of camera

 

[Bladerskb]

I don't fault Tesla for using map data. But it is funny how Tesla fans say that HD maps are a crutch. Yet, Tesla uses map data.

And not just use it, its vitally critical. No map? It won't stop at the intersection.
There is a map but no light? It will still stop at the intersection.

Which is why they said it won't work with temporary lights or intersection.

But don't worry, the tesla people will revise history to reshape their narrative and make this totally okay.

Example:

If you need maps for anything other than finding a route to your destination, it's not generalized.

 

[willow_hiller]

I wonder why they don't have it beep or vibrate the steering wheel to ask for confirmation? Then allow a confirm with a tap of accelerator that would not cause the car to accelerate.

I thought this is how it is going to work? Just with the screen notification instead of a vibration. If you watch the video in the tweet, the speedometer never decreases while the car is asking for input before a green light. As long as the driver gives acknowledgement 200 feet away from the intersection, the car doesn't slow down at all for green lights.

 

[ZeApelido]

And not just use it, its vitally critical. No map? It won't stop at the intersection.
There is a map but no light? It will still stop at the intersection.

Which is why they said it won't work with temporary lights or intersection.

But don't worry, the tesla people will revise history to reshape their narrative and make this totally okay.

Example:

It is pretty embarrassing, even if only "temporary"

 

[diplomat33]

And not just use it, its vitally critical. No map? It won't stop at the intersection.
There is a map but no light? It will still stop at the intersection.

Which is why they said it won't work with temporary lights or intersection.

But don't worry, the tesla people will revise history to reshape their narrative and make this totally okay.

Example:

Agree. And Tesla itself puts in the manual that traffic light response may not be reliable in these conditions:

- Visibility is poor (heavy rain, snow, etc) or weather conditions are interfering with camera or sensor operation.
- Bright light (such as direct sunlight) is interfering with the view of camera(s)
- A camera is obstructed, covered, damaged or not properly calibrated.
- Driving on a hill or on a road that has a sharp curves on which the cameras are unable to see upcoming traffic lights or stop signs.
- A traffic light, stop sign, or road marking is obstructed (for example, a tree, a large vehicle etc)
- Model 3/Y is being driven very close to a vehicle in front of it, which is blocking the view of camera

And I noted in another post that all these limitations would probably be solved with extra sensors and/or HD mapping. These limitations admit that there are situations where cameras alone are not sufficient. So really, the need for HD mapping should be obvious since there are cases where cameras alone won't be enough like when a traffic light or stop sign is obstructed. You need HD mapping to tell the car about these objects that your sensors can't detect.

 

[Sharps97]

No chance in hell I’d run something with a list of warnings that long. Just not even going to risk it, so no data from me. Takes too long to go through the parts replacement process to risk the damage this will cause. And since Tesla assumes no liability, sorry but no thanks I’m not a multi billion dollar business. Honestly, Tesla should be embarrassed to release something like that.

 

[Daniel in SD]

I thought this is how it is going to work? Just with the screen notification instead of a vibration. If you watch the video in the tweet, the speedometer never decreases while the car is asking for input before a green light. As long as the driver gives acknowledgement 200 feet away from the intersection, the car doesn't slow down at all for green lights.

Ah. I didn't read the Note! "at any time after touchscreen displays the read stop line"
I never understand how people are able to drive this car and look at the screen at the same time. When I'm approaching a light I'm not looking down at the screen! Seems like a beep or steering wheel vibrate should be an option.

 

[Bigkidatheart]

Hi, all. I got a notice when I started up my Model X that "Full self-driving update is complete.". However, I can't see anywhere where I can activate this. On normal autopilot and Autosteer, I don't notice it doing anything different than it has before. I just got this notice maybe two days ago. Has anyone else gotten this notification? And do you know what is needed to activate this?