-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla autopilot HW3
- Thread starter verygreen
- Start date
I may be misunderstanding you, if so disregard...
The NN runs in O(n) time, all fixed operations, no conditionals, so it does not (or at least should not) randomly drop frames. The frame rate and image size are chosen to put the computation within the frame period.
Q3 2018 conference call transcript, available at Motley Fool. HW3 is not yet running the new NN on public cars, but it is larger than what HW2.x can support.
I don't think AP1 has an upgrade path. The hardware set is too different. Tesla is talking about changing the CPU board and not all of the sensors.
As an aside, I still wonder why the split with MobilEye happened. I have 2.5 and it doesn't recognize speed limit signs like a friend's AP1 car.
Thanks! How can they cram stop light recognition if it's already maxed out? Then something has to give, right?
Also I read that the raindrop detection took like 30% power, can't remember where.
I think ME was protecting their IP by not letting Tesla train their NN in parallel with it (free labeling).
I didn't say 2.x was maxed, just that the new 3.0 intent NN is much larger. Tesla/ Karpathy and team are adding as much functionality to HW2.x NN as possible. Under the SW 2.0 development method, to add a feature they add in training data and test cases along with the NN size/ shape and run the training. If the new (2.x) NN hits the performance metrics, the new functionality can stay, if failure rates increase, then something must be tweaked, or else no new feature.
1. I intuit it from the fact that every time someone is killed using AP, I never hear Musk discuss how its inability to detect massive stationary obstacles in the planned path at highway speeds will ever be robustly solved, instead he performs a ritual shifting of all blame onto the deceased and tries to fob us off by regurgitating more statistical happy-clap-trap.
2. I do not dispute that inattentive drivers contribute to their own deaths, but that is beside the point, which is how to minimise the contribution of AP/FSD?
3. Building further complexity [FSD] upon the IMHO unsound basis of AP sensor package, which will moreover be activated in the city amongst vulnerable pedestrians and cyclists, indicates that he is willing to incur an increased rate of fatalities/injuries rather than fix the fundamental lack of redundant sensors which are reliable at all operable speeds.
4. Yes, by my count AP contributing to 5 fatalities in 38 months is too many, too regularly ... 1 case every 30 weeks. A feeling exacerbated by the fact that there is no plan to reduce it in a safe and reliable way.
Last edited:
That's a lot stated as fact. How do you know what Musk feels? How do you know he is cavalier about safety? I have NEVER hear any rationalization as you mention. I do believe that Musk believes AP is safer overall. BTW have you ever thought that we rationalize safety and deaths all the time? Put speed limiters in cars so they can't go over 15 mph. I bet US traffic deaths would drop by 25,000 per year or more. But, we like to get to work sooner and take faster trips so we tolerate the deaths. By comparison you are talking a small number of deaths where the people weren't paying attention. What about the times AP avoided an accident?
I'm curious, why all the Tesla hatred? Did you have a lot of trouble with your car? Your post drips hatred. I am curious as to the source of that hatred.
I know you didn’t ask me but if it helps: I find Tesla as a company has been disappointing in many ways, leadership included. The car itself is good. Also AP2 as a product has been disappointing.
Bet TSLA
Active Member
You have some special insight into how many fatalities AP has prevented? I think even the relatively simple superhuman ability to see an extra car ahead has saved way more lives and vast quantities of property. The car brakes for no apparent reason because the car in front is closing rapidly with the car in front of it. This means that you and the the car behind you and the cars behind them *do not* get involved in the mess, whatever it turns out to be.
This is just one of several ways that the existing autopilot software has repeatedly prevented fatalities.
1. I can't actually know how he feels but am discussing the trade-offs he is obviously willing to make over a period of 3 years since the first fatality to which AP contributed. From this I can deduce that saving costs is more important to him than wasting a few more customers' lives.
2. He says AP is safer overall, because, well, he needs to keep selling it, but neither you nor I know whether he truly believes that. Certainly the stats he uses are pretty questionable.
3. I conclude he is cavalier about safety because he has failed to resolve a fundamental safety problem [AP ploughing into massive obstacles in planned path at 80mph without brakes or warning tone] yet now proposes to build much more complex functionality, likely to induce an even greater degree of driver complacency, upon this unsafe sensor basis, namely the inadequate radar and driver attentiveness monitoring system.
4. No, I've had relatively few problems with my car. Further I do not hate Tesla in the slightest and support the declared mission with much monies. What I am sceptical about is how the car I have paid for will ever reach the promised FSD in a manner which is safe and reliable, such that I could trust it to drive unsupervised at L3 without having to worry about my wife & children ending up plastered @80mph into the tail-end of the first stationary traffic jam it encounters?
5. IMHO the only way that can ever happen is with, at minimum, a radar sensor upgrade to go along with HW3. And a DAMS worth a damn, i.e. which tracks that the driver's eyes are actually open and focussed on the road ahead, would also be a great help in getting to L3 approval without unnecessarily wasting further customers/3rd parties.
1. No, I have no special insight on that but of course accept that in dynamic situations AP works well to accurately detect vehicle motion ahead and thus prevents a good many crashes, which are perhaps in the main low speed fender-benders in dense traffic.
2. That laudable fact however does not compensate for the treacherous weakness that at any moment doing 80mph on the highway it is liable to go full tilt into a massive obstacle parked in the planned path, the type of accident on restricted-access highway with the highest probability of resulting death.
3. It has neither been explained by Tesla why this weakness persists after 3 years, nor how they plan to eliminate it before building FSD upon the same sensor basis.
@OPRCE I think the simple reality is, Tesla has found it much harder to do reliable vision than they (or Elon) thought back in 2016.
I think I actually said that, though it was meant as a sarcastic joke! Probably should have marked with /s, sorry.
In fact I would imagine it consumes only about 5% of processor power, which is still a terrible waste.
lunitiks
Cool James & Black Teacher
Only the cameras can determine if they are occluded, a rain sensor at a different location provides zero usable data regarding bird droppings in the field of view...
Ok, that is just not true. Why are you only looking at cherrypicked Tesla incidents, when so many people die in car accidents every day?? Society puts up with those "human sacrifices" because the alternative would be to just stay in one place. But Tesla is one of the leading companies trying to stop all those deaths.
Nearly 1.3 million people die in road crashes each year, on average 3,287 deaths a day. An additional 20-50 million are injured or disabled. More than half of all road traffic deaths occur among young adults ages 15-44.
Sure, it would be ideal to have LIDAR and more sensors, but then nobody could afford it. FSD is already saving lives; let's face it, improving on typical human drivers is not hard to do.
Well, it could in theory do that for the forward three at least...
But again, in practice it does not work.
There is no cherrypicking, I am looking at all deaths to which AP contributed, which AFAICT = 5 cases.
The statistics on deaths/injuries in other vehicles are immaterial, AP remains fundamentally unsafe at higher speeds and Tesla has shown no plan to make it safe. Even if tomorrow HW3 & FSD NNs rendered the vision system 99.99% accurate, in the case where the system fails to recognise from camera input that it is moving at 80mph into a static solid obstacle, the radar still provides no redundancy and another fatal crash can ensue if driver inattentive in the crucial moment. Then for actual L3 no driver attentiveness is required, making the fatal crash inevitable.
If the radar were upgraded to something fit for purpose, i.e. capable of reliably distinguishing stopped objects when moving at up to 90mph, L3 SAE may well be achievable without LiDAR. Certainly it is worth trying for.
Last edited:
Similar threads
- Article
