Both are launching a commercial taxi service. Waymo is already running, just got a new licence for California in fact.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Elon: "Feature complete for full self driving this year"
- Thread starter dckiwi
- Start date
Both are launching a commercial taxi service. Waymo is already running, just got a new licence for California in fact.
I get that but aren't Waymo limited to certain locales, say San Francisco vs Providence RI, whereas a tesla, not requiring a premapped area could, in theory, navigate from San Francisco TO Providence RI. It's a different problem space as one simulates a human driver's perception of where they are whereas Waymo requires a pre-established hyper accurate mapping of the area.
boonedocks
MS LR Blk/Blk 19”
if you are taking Tweets / public statements / etc etc etc into account then YES Elon has promised that...and much much earlier than out of the gate this year. SMH
Yah general solution. Waymo also does not do parking lots (that I am aware of).
So what is more difficult?
Explaining gravity, or what Elon meant by that?
After reading this, I wondered, if I spent as much time trying to explain gravity
as I have reading threads about what Elon meant by this or that
I could probably explain grav...
Nah. I still woudn't have the answer
To either question.
: )
Explaining gravity, or what Elon meant by that?
After reading this, I wondered, if I spent as much time trying to explain gravity
as I have reading threads about what Elon meant by this or that
I could probably explain grav...
Nah. I still woudn't have the answer
To either question.
: )
Daniel in SD
(supervised)
They do:
Waymo CEO John Krafcik explains why a parking lot is one of the most difficult environments for a self-driving car
Though this customer says it failed in the Costco parking lot. Not surprising...
Waymo Self-Driving Cars Are Getting Confused by Rain, Passenger Says
diplomat33
Average guy who loves autonomous vehicles
Then how do you explain that Elon also said feature complete would require driver supervision and robotaxis would be end of 2020? If it requires driver supervision until sometime in 2020 and robotaxis are end of 2020, FSD can't be L5 with no supervision at the end of 2019.
Last edited:
The feature-set allowing for the car to drive itself is a practical matter governed by the accomplishments of Tesla. Robotaxis and a car being able to drive around without supervision is a regulatory matter governed by what the US government deems an acceptable safety level. The two are totally different, despite the latter requiring the review of the former.
@diplomat33
This is all very simple really, in my view you are making it unnecessarily complex.
Either Tesla have implemented the features to be Level 5 at the end of 2019 or they have not.
They do not need the regulatory approval or the reliability to drive it driverless, just the implementation, to be feature complete.
If they are missing major features (like missing speed sign recognition entirely or somesuch), then they can not be Level 5 feature complete.
Very simple really.
This is all very simple really, in my view you are making it unnecessarily complex.
Either Tesla have implemented the features to be Level 5 at the end of 2019 or they have not.
They do not need the regulatory approval or the reliability to drive it driverless, just the implementation, to be feature complete.
If they are missing major features (like missing speed sign recognition entirely or somesuch), then they can not be Level 5 feature complete.
Very simple really.
What if the required supervision is regulatory required pending better real miles test data? One way to explain it is supervision is a separate thing from is the software different from a fully/better trained network. In other words, what is released end of 2019 is capable of full level 5 but for better training set and regulatory approval. This gets back to the definition of feature complete.
see Train your first neural network: basic classification | TensorFlow Core | TensorFlow
In other words, the software is ready, the dataset isn't.
They need reliability in order to get regulatory approval to remove the requirement to have a driver holding the steering wheel.
The size of their training base determines, in part, the accuracy of detection 99/100 or 50/100. If it 'gets it right' 50% of the time nobody is going to approve it for driverless operation. The software is fine, the training database is insufficient.
In this scenario the software is L5 ready but the training isn't. So come, according to Musk, mid 2020 the training will be sufficient and it gets downloaded, the requirement for holding the steering wheel gets removed, references to Beta get removed and off you go.This IS pretty simple.
I believe you are confusing training set and validation dataset. The software relies on training dataset to function. Validation dataset is when you prove that it does what it should when it should.
I would accept that scenario as feature complete Level 5 no geofence at the end of 2019, if that is where things really stand with Tesla’s software then.
boonedocks
MS LR Blk/Blk 19”
My post wasn’t referring to you leaving out a part of Elon’s promises.....It was actually that you are referring to the rest of your post as promises by Elon.
I do find it interesting at how charmed you are with Elon...even to the point of, I believe, would even defend him if he was completely indefensible.
A lot of us on the forum were once smitten as you are and would counter any logical explanation or example with rose colored glasses. It’s just that most of us that were that way have had our eyes opened and don’t believe anything until it is actually installed or updated on our cars. I don’t say that from a jaded perspective but from a “been there done that” reality point of view. I do believe that the probability of that will be pretty great with you too.
In the last day to two Elon has publicly stated that HW 2.0 and 2.5 should start being retrofitted with HW3.0 the 4th quarter. This is an example of something that has caused a quite a few to start shouting it from the rooftops. Do I believe it will actually happen? Nope I do not. Hopefully Ill have to come back at the end of December and eat my proverbial “hat” I would guess that chance in the low single digits as there may be a car to two updated just to proclaim that it was done. ¯\_(ツ)_/¯
I am not a shorter or a Fudster as I just had my 4th Tesla drivered Mid-June. I have however been around the block with Tesla a time or two and see what comes from Elon as wishful and hopeful thinking and not at an outlook or even probability of may or may not happen. I do believe you too will one day clean the rose color from you glasses and just defend that things that are reality. This is not.
Last edited:
Well, over the last several days of discussing this with you, absolutely, what Musk is saying, if found to be totally factually incorrect, then I agree with you, it's a whole 'nother ballgame and SEC will be involved etc. So, in my mind is there a middle ground where what Musk says is accurate given the best information available to him at the time. If that's the case what would explain feature complete (a software term) and NOT ready for driverless use (yet). My argument is something non-software related. If, come 12/31/2019 I cannot tell my car to drive to Stop&Shop even with my hand on the wheel but, on 1/31/2020 I can, is that an abysmal failure on Musk's part or is that a function of imprecise science. He, rightfully should get some modicum of wiggle room, just not measured in years. I say that as I've heard two things, driverless by mid 2020 and end of 2020. What if the truth is Sept 2020 people can elect to have their steering wheels removed and steering wheel is an option when ordering. But sometime between end of this year and very early next I would hope, and expect, we can have our cars drive us anywhere with nag requirement but have a self driving reliability better than 95% of the time.
The next bitter battle will likely be over better driving than the avg human and who is avg. I suspect if that alleged drunk in the pickup were in a Tesla on AS or NoA he never would have sideswiped that Tesla. Does avg human include those texting, talking, leaving a bar for home on Fri night? I suspect we all know we will hear, "had I been driving that never would have happened". One cannot prove a negative.
Suffice it to say, I expect to be able to tell my car where to take me by end of 2019 or VERY soon thereafter. TWT.
diplomat33
Average guy who loves autonomous vehicles
You are mistaken. I am not charmed. I've been clear and consistent lately that I doubt Tesla will get to L5 on Elon's timeline.
I am a realistic the most I learn about self-driving. Feature complete will not be L5 by the end of the year. So if you think that by Dec, Tesla will release L5 city self-driving without supervision, you are going to be sorely crushed.
diplomat33
Average guy who loves autonomous vehicles
And my stance is that Tesla's feature complete will not be L5, period! It will be designed with the same ODD as L5 but it will NOT be L5. That is what I've been trying to tell you guys!
You can argue "But Elon said L5" all you want but I am telling you reality. FC will not be L5 by the end of this year!
And that could well be the case. I bought a, presumably, MS level book, "Neural Networks and Deep Learning" by Aggarwal which has been, to be kind, a challenge. TensorFlow tutorial, way more digestible. I would love someone with deep Neural Network / Deep Learning expertise to join in this conversation. My machine learning familiarity (won't even say expertise) ends with Mahout. The issue I think people are grappling with on here, people with little to no software expertise, little to no engineering expertise, little to no machine learning expertise is can the software be solid but the training (/ validating) data be insufficient to reach the reliability necessary for self driving. My experience in software instructs me if they say, by year's end, the software necessary for L5 compliance will be feature complete, they should know if that is a true statement plus or minus some modicum of wiggle room. In other words if not released by 12/31 but released 2 weeks later. I've never seen a ship date slide without dire consequences. That's what nights and weekends are for. Especially in a publicly owned company, thou will make your dates.
This is an over simplification, but still an excellent video:
And the followup is also good:
The tweaking of the system in Tesla's case is largely (but not entirely) the feedback drivers give when they take the wheel back from autopilot. There's also a lot of manual data labeling that was alluded to in autonomy day.
Vice news did a good piece on this here:
Once Tesla is reasonably confident that the shadow mode driving NN doesn't differ from the way the real world drivers are driving (validating the system that had been trained by previous data inputs, Validation Mock 1) then they probably upgrade it from development to production. Once its in production every mile that autopilot drives without drivers disengaging it equates to additional validation (given certain inputs the NN makes a driving output that you as a driver by not disagreeing with the choice of have effectively validated, Validation Mock 2). At some point Tesla goes to DC and says "hey look at this validation, people don't have to drive the car themselves ever". Then DC people will go, Ok then, you can turn off the need to hold the wheel and we will let the empty car drive itself.
Last edited:
Similar threads
