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

[strangecosmos]

They capture, but aren't triggered by what the driver is doing.

All that’s needed for supervised learning of path planning and driving policy (a.k.a. behavioural cloning) is the state-action pairs. State = what the car observes. Action = the driver input.

The state-action pairs could be collected completely at random, or when certain hand-coded conditions are met. It doesn’t matter. All that’s required is that the state-action pairs are uploaded.

Also, the state doesn’t have to be the raw sensor data. It can, in theory, be the set of labels, bounding boxes, spatial information, etc. outputted by the neural network.

 

[eli_]

My best guess prediction is what they'll ship this year is extending Navigate on AP to handle additional maneuvers on local streets, i.e. more turn-by-turn nav steps will gain the "blue wheel" icon indicating AP can handle that step. That will include some partial set of intersections (more complex intersections will be unsupported in the map data), and most right turns, but unprotected left turns will not be supported like Waymo, they're just too dangerous. The car may route around them or just disengage AP and require manual control.

I think it'll make a cool demo but it'll be very stressful for users, it will be much easier to just drive the car yourself, and you'll probably reach the destination much faster. Nav on AP is already a bit stressful to monitor even on closed highways, when you're babysitting the car in urban environments through intersections it'll be like a full time job. I mean they literally train and pay people to do that sort of thing.

 

[rnortman]

I think it'll make a cool demo but it'll be very stressful for users, it will be much easier to just drive the car yourself, and you'll probably reach the destination much faster. Nav on AP is already a bit stressful to monitor even on closed highways, when you're babysitting the car in urban environments through intersections it'll be like a full time job.

This is really the whole point. Yes, Tesla will eventually deliver something which they call "FSD". And just like the existing NoA feature, it will be nothing but a gimmick. It will not improve your safety nor reduce your stress; rather it will likely do harm on both fronts if you choose to make use of it. And it will never ever be L5 as Elon originally promised; it wil be a miracle if they manage to get to L3 on the highway; local/urban environments will remain L2 forever.

 

[BestRadar]

This is really the whole point. Yes, Tesla will eventually deliver something which they call "FSD". And just like the existing NoA feature, it will be nothing but a gimmick. It will not improve your safety nor reduce your stress; rather it will likely do harm on both fronts if you choose to make use of it. And it will never ever be L5 as Elon originally promised; it wil be a miracle if they manage to get to L3 on the highway; local/urban environments will remain L2 forever.

I bought FSD because of the conveniences they claimed in the video. I want to pull up to the office and have my car go park itself and then come pick me back up at the door when I am ready.

Lets not forget their exact description which was on the Tesla website when I purchased the feature in 2016

"All you will need to do is get in and tell your car where to go. If you don’t say anything, the car will look at your calendar and take you there as the assumed destination or just home if nothing is on the calendar. Your Tesla will figure out the optimal route, navigate urban streets (even without lane markings), manage complex intersections with traffic lights, stop signs and roundabouts, and handle densely packed freeways with cars moving at high speed. When you arrive at your destination, simply step out at the entrance and your car will enter park seek mode, automatically search for a spot and park itself. A tap on your phone summons it back to you."

 

[rnortman]

I bought FSD because of the conveniences they claimed in the video. I want to pull up to the office and have my car go park itself and then come pick me back up at the door when I am ready.

You and everybody else who purchased based on that promise deserve nothing less. But Tesla will not be able to deliver it. The only question is what, if any, compensation you will receive.

 

[Soda Popinski]

This is really the whole point. Yes, Tesla will eventually deliver something which they call "FSD". And just like the existing NoA feature, it will be nothing but a gimmick. It will not improve your safety nor reduce your stress

That's a bit silly. EAP already reduces my stress on my freeway commute. How can an improvement not be less stressful? Even if we don't see "FSD" on surface streets, I'd expect freeway "FSD" to get better than EAP, not worse.

My best guess prediction is what they'll ship this year is extending Navigate on AP to handle additional maneuvers on local streets, i.e. more turn-by-turn nav steps will gain the "blue wheel" icon indicating AP can handle that step. That will include some partial set of intersections (more complex intersections will be unsupported in the map data), and most right turns, but unprotected left turns will not be supported like Waymo, they're just too dangerous. The car may route around them or just disengage AP and require manual control.

Don't expect it to ship to owners this year. Musk specified "feature complete" at the end of the year. Even if we go by that optimistic timeline, feature complete is basically the start of "beta." All the features are in, instead of placeholders, but will be expected to be unreliable until the bugs are worked out. Musk gave a timeline for that of end of 2020.

 

[im.thatoneguy]

How can an improvement not be less stressful?

Because it'll increase the domain of time that you are frustrated by substandard behavior. Have you used NoA? EAP isn't terribly stressful, you turn it on... if you want to change lanes you disengage, pass and re-engage. NoA for me is just a ton of mental work because not only am I deciding when I need to change lanes it is too and I'm having to negotiate which ones it are suggesting are real and which are completely wrong.

For example yesterday EAP said pick any of the 4 lanes to proceed. But only 2 of the 4 actually would take me to my destination and if I waited until EAP said to change lanes, I let alone it would never make it over across 2 lanes in time. So not only are you stressed about making lane changes, but you have to also add the stress of rejecting someone's very wrong opinion which makes you second guess yourself. "Take the left lane." "I'm pretty sure the right lane is the one that goes through." "Change lanes to the left!" "No! go to the right! Why aren't you going to the right. Oh god, did you just swerve back into the left lane?! I said change lanes to the right that was our gap, I give up let me drive."

EAP used to only work on controlled highways then they let us use it on surface streets. It works fine except it has to swerve over into merging lanes. So now instead of using human autopilot I have to anticipate when it's going to merge, disengage and re-engage every time a merge lane comes up. Things that the human brain can just do on its own autopilot require vigilance and intervention. Or this morning I had to anticipate autopilot stupidity like slamming on the brakes when a car makes an unprotected left in front of me, even though they are well beyond my stopping distance and decision point for an emergency maneuver. So not only do I have the stress of judging a car passing in front, I also have the stress of judging autopilot's stupidity level to counteract its dangerous reaction.

 

[Soda Popinski]

Because it'll increase the domain of time that you are frustrated by substandard behavior. Have you used NoA? EAP isn't terribly stressful, you turn it on... if you want to change lanes you disengage, pass and re-engage. NoA for me is just a ton of mental work because not only am I deciding when I need to change lanes it is too and I'm having to negotiate which ones it are suggesting are real and which are completely wrong.

For example yesterday EAP said pick any of the 4 lanes to proceed. But only 2 of the 4 actually would take me to my destination and if I waited until EAP said to change lanes, I let alone it would never make it over across 2 lanes in time. So not only are you stressed about making lane changes, but you have to also add the stress of rejecting someone's very wrong opinion which makes you second guess yourself. "Take the left lane." "I'm pretty sure the right lane is the one that goes through." "Change lanes to the left!" "No! go to the right! Why aren't you going to the right. Oh god, did you just swerve back into the left lane?! I said change lanes to the right that was our gap, I give up let me drive."

EAP used to only work on controlled highways then they let us use it on surface streets. It works fine except it has to swerve over into merging lanes. So now instead of using human autopilot I have to anticipate when it's going to merge, disengage and re-engage every time a merge lane comes up. Things that the human brain can just do on its own autopilot require vigilance and intervention. Or this morning I had to anticipate autopilot stupidity like slamming on the brakes when a car makes an unprotected left in front of me, even though they are well beyond my stopping distance and decision point for an emergency maneuver. So not only do I have the stress of judging a car passing in front, I also have the stress of judging autopilot's stupidity level to counteract its dangerous reaction.

Ok. That explains why current AP and Nav on AP is stressful. Why would an improvement make it more stressful? EAP has improved since I got the car in July, decreasing my stress while driving. NoAP is still pretty crappy for heavy stop and go, so I don't use it on my daily commute. I'd imagine "improvements" will make it better, and hopefully usable. Until then, I don't use it, so I don't stress.

 

[mongo]

Don't expect it to ship to owners this year. Musk specified "feature complete" at the end of the year. Even if we go by that optimistic timeline, feature complete is basically the start of "beta." All the features are in, instead of placeholders, but will be expected to be unreliable until the bugs are worked out. Musk gave a timeline for that of end of 2020.

in the ARK podcast, he indicated driver supervised this year and sleep while riding in 2020.

 

[Soda Popinski]

in the ARK podcast, he indicated driver supervised this year and sleep while riding in 2020.

I took that to mean employee test driver, not consumers, but you could be right. My reasoning is because I've seen "feature complete" software, which were so buggy at that early beta stage, to not be anywhere near ready for consumers. Secondly, he mentioned feature complete by the end of the year, which is after driver supervised. So that would mean driver supervised feature incomplete "FSD," which sounds like either piecemeal aspects of FSD (say hands free freeway driving), or Tesla employee test drivers running FSD with placeholders, or both.

 

[mongo]

I took that to mean employee test driver, not consumers, but you could be right. My reasoning is because I've seen "feature complete" software, which were so buggy at that early beta stage, to not be anywhere near ready for consumers. Secondly, he mentioned feature complete by the end of the year, which is after driver supervised. So that would mean driver supervised feature incomplete "FSD," which sounds like either piecemeal aspects of FSD (say hands free freeway driving), or Tesla employee test drivers running FSD with placeholders, or both.

Yah, I hear ya. Given the recent internal Tesla test fleet expansions, I'm leaning toward consumer release this year. My thinking is that he meant public level feature complete. If it is merely an internal benchmark, there was no reason to mention it to the public with the caveat of needing monitoring. So all FSD functions with EAP nags.

 

[lunitiks]

in the ARK podcast, he indicated driver supervised this year and sleep while riding in 2020.

Elon suggests his software (and hardware) is Level 4 ready next year?

Seriously, his timelines have gone from exciting to unreasonable to just sad.

I really think it’s reached the point of sad now

 

[mongo]

Elon suggests his software (and hardware) is Level 4 ready next year?

Seriously, his timelines have gone from exciting to unreasonable to just sad.

I really think it’s reached the point of sad now

You would prefer his timelines get further out instead of closer?

 

[BestRadar]

You would prefer his timelines get further out instead of closer?

Yeah, I would like to see reasonable expectations and some apologies for the past exaggerations.

 

[mongo]

Yeah, I would like to see reasonable expectations and some apologies for the past exaggerations.

Yeah, two years seems to be his limit on how long something will take.

Estimates are hard with software. Estimates are even harder when dealing with a team of coders. The way Elon codes is to go as close to 24/7 as he can till it's done. That doesn't work if you want to retain employees. He also had some hiccups both external (MobilEye breakup) and internal (3? Autopilot leads). However, the recent events and guidance make me hopeful that there is real data/ progress to support the latest estimates (even if the long term one is still bumping against the 2 year boundary).

 

[BestRadar]

Yeah, two years seems to be his limit on how long something will take.

Estimates are hard with software. Estimates are even harder when dealing with a team of coders. The way Elon codes is to go as close to 24/7 as he can till it's done. That doesn't work if you want to retain employees. He also had some hiccups both external (MobilEye breakup) and internal (3? Autopilot leads). However, the recent events and guidance make me hopeful that there is real data/ progress to support the latest estimates (even if the long term one is still bumping against the 2 year boundary).

My problem is they released a video that claimed to be a trial run on existing software and hardware back in 2016 when I jumped on board. Nowhere did they claim they needed 2+ years of engineering when they were selling me on the FSD product. They claimed the only delay was software validation and regulatory approval.

 

[mongo]

My problem is they released a video that claimed to be a trial run on existing software and hardware back in 2016 when I jumped on board. Nowhere did they claim they needed 2+ years of engineering when they were selling me on the FSD product. They claimed the only delay was software validation and regulatory approval.

Yeah, I can understand the frustration.

 

[rnortman]

That's a bit silly. EAP already reduces my stress on my freeway commute. How can an improvement not be less stressful? Even if we don't see "FSD" on surface streets, I'd expect freeway "FSD" to get better than EAP, not worse.

When it tries to do more than it can, it gets worse. EAP is great most of the time. NoA is terrible, because it tries to do more than it can. When it's trying to handle traffic lights it will get it wrong frequently enough that it will be more trouble than it's worth.

 

[rnortman]

Estimates are hard with software.

Normal software projects are hard to estimate. This isn't just a normal software project. This is more like basic research. There's not an engineer on the planet who can outline in a known/proven way to achieve what Elon has promised they will achieve -- by which I specifically mean vision-only (well, OK one radar too) L4/L5 autonomy.

 

[eli_]

When it tries to do more than it can, it gets worse. EAP is great most of the time. NoA is terrible, because it tries to do more than it can. When it's trying to handle traffic lights it will get it wrong frequently enough that it will be more trouble than it's worth.

Yeah basic auto-steer/lane keeping isn't very stressful, because the car is just staying within the lines, keeping a speed. It's not moving laterally or making executive decisions and planning things, and you just need to keep your eyes straight ahead to make sure it's not wandering out of the lane.

When you add in NoA with the addition of automatic lateral movement to change lanes or take ramps and interchanges, handle merges, and then eventually local intersections with traffic lights or stop signs, it expands the range of things you need to monitor. You need to keep situational awareness all around the car, plus monitoring the screen for what the car is planning to do, plus babysitting a beta robot driver all at the same time. The mental load feels higher than just driving, but like anything it can probably get easier with practice.