I guess I get confused with the definition of Full.
Maybe it should be SSD - Semi Self Driving
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
I guess I get confused with the definition of Full.
Maybe it should be SSD - Semi Self Driving
JeffK
Well-Known Member
We are talking about full self driving with different tasks handled by different networks instead of one massive network. One massive network would be too generalized to handle corner cases properly but would be just fine for simple demos. IE. Nvidia's BB-8 demo
Kant.Ing
Member
About corner cases, they desperately want near death or near accident events, otherwise the 6 billions millions are useless. So, drive harder!
https://www.linkedin.com/pulse/due-diligence-recommendations-mobile-autonomous-industry-dekort
https://www.linkedin.com/pulse/due-diligence-recommendations-mobile-autonomous-industry-dekort
Reciprocity
Active Member
Yes, but there are enough corner cases to resolve that any driving will help the machine learn. The siutations you are talking about might need to be manually guided through the system by a developer. The goal is to limit how many situations need human intervention to provide a solution. Its almost like trying to teach your teenager to drive and always taking over when you see an issue coming up. The machine needs to experience and properly resolve as many situations as possible.
Reciprocity
Active Member
The pothole is the easiest as long as some people are manually avoiding them. The system should learn that there is an issue there and over time, learn what a pothole looks like and how to avoid them. The question is what does the car do when the pothole is fixed. It should learn that people stop avoiding a specific location and learn to stop avoiding it. I don't know if potholes will be uniform enough to truly identify them. I think with enough autonomous cars on the road, you will see systems, some software and some manual, that will mark these things. In theory you could paint a reflective circle of paint around the pothole with an X through it so that the car could see it better and know it is something to avoid. This could be physically painted on or just tagged in the high def map. Once the location is repaired, the tag/paint is removed.
Passing a stationary car is no different then passing a double-parked car. The Bolt/Cruise Automation video that was recently on Electrek, shows that happening successfully today. Same with pedestrians on bikes if you watch the entire video and even slows for a small animal darting in front of the car.
You are not supposed to swerve to miss animals anyway as it could cause an accident. My guess is that over time, the machine learning process will start to pickup that people driving tend to slow and drive around these mounds of redish lumps and will learn to do the same as long as the way is clear. This is one of the more complex issues as there could be any number of types of road debris that would be harder to identify. In theory, if the system identifies something in time, it could compare it to recent images of the location. The issue is that the high def maps that would be stored locally probably wouldn't have this type of detail and I don't know if it could request this info in time to compare it as you drive. More then likely, the thing the first car sees that it cant figure out would be sent to the mothership to analyze and by the 3rd or 4th car, the system would know whether it should be avoided. Some of those cars will be manually driven and some autonomously. i could also see data sharing among autonomous software providers and certainly real time info shared between cars that are close by and both of which might be required by law. This would speed up the identification of such issues dramatically and probably more people will hit the object driving manually then autonomously.
Reciprocity
Active Member
I am sorry, but I just read this article and it full of completely useless, asinine and nonsensical arguments as well as being chock full of false or misleading assertions. If people like this had their way, 200,000 people would continue to die on the roads to accidents that could have been avoided by driver assist solutions like AP1/AP2 and even more lives saved with FSD. We cant allow perfect to be the enemy of good in this situation and Tesla has the right approach of getting the tech into it's "guinea pigs" hands asap so it can save as many of their lives as possible. The one death that had any AP related issues was the driver who drove directly into the side of a big rig at full speed without breaking. This clearly shows that the driver was not paying attention at all. Yes the system failed in that case, but how many times on that same trip did it save that persons life even though they were using the system inappropriately?
Autonomy is coming. Its coming sooner then most think and driving a car will be outlawed in many places before you know it. And you wont care because the convenience and expense will be so beneficial that you wont be able to imagine a world where people need to own a car. I am talking 10-20 years and most car ownership for people will end. Some will still own hotrods and collector cars, but driving them will be highly restricted to certain places. It will be scary to drive on some roads because the autonomous cars will be going 2x the traditional speed with 1/2 the space between each car. Though you will still be able to drive right down the middle of them and even go the wrong way on a one way and they will avoid you and communicate that a human driver has gotten on the road again and route all the traffic around the vehicle to allow it to get clear of the area.
stopcrazypp
Well-Known Member
This is the same guy who wrote this thread (perhaps the most disagreed thread in TMC history):
Tesla's Autopilot needs to be shut down and NHTSA needs to do their due diligence
The article above uses the exact same key words (Commercial IT, Scenario Matrices, exception handling, NHTSA due diligence, regression). The thread clearly demonstrates he doesn't know what he is talking about in regards to autopilot (simple example: he appears to not even know AP1 and AP2 uses different hardware until someone kindly pointed it out). See the thread for more details.
Last edited:
Its a myth that tesla autopilot doesn't have any competition and this is why Tesla continues to be a breeding ground for myths.
For example 2017 Volvo Pilot Assist 2 (It even handles curves better than AP1 and AP2). There's no dumb braking and not recognizing car infront, etc... Obviously Volvo doesn't have the cult following that Tesla has so there are no thousands of video of their pilot assist 2 randomly driving around.
rain/snow/sleet
local road
huge bend
JeffK
Well-Known Member
No one said Tesla didn't have competition with regards to autonomous driving.
If you're such a big fan of Volvo / MobilEye then it's settled, go buy an EV from Chinese owned Volvo.
They even have a forum site you can troll Volvo Forums - Volvo Enthusiasts Forum
Here's a pic of their EV concept:
Comes with it's autopilot Max Headroom:
Also most people don't know that Volvo has OTA and believes Tesla are the only one. Another myth.
Here is a Volvo 2016 model that has been updated from Pilot Assist 1 to 2, including map update OTA.
Pilot Assist 1 was limited to 30 mph and required a lead car, while Pilot Assist 2 is limited to 80 mph and requires no lead car.
Here is a Volvo 2016 with a software update running pilot assist 2
New Features available as of November 2016
Here is a Volvo 2016 model that has been updated from Pilot Assist 1 to 2, including map update OTA.
Pilot Assist 1 was limited to 30 mph and required a lead car, while Pilot Assist 2 is limited to 80 mph and requires no lead car.
Here is a Volvo 2016 with a software update running pilot assist 2
New Features available as of November 2016
Last edited:
Reciprocity
Active Member
My AP2 handles all those situations with no issues. I know some have been reported, but my guess is that those folks need cameras adjusted it something. Rain, corners even sharp corners like circular exit ramps at speeds that I would do manually.
Volvo is a solid, safe car. I commend them. But AP2 or AP1 is not FSD1.0. they are different systems just like Volvo's solution undoubtedly is. Tesla leads in the fact that they have cars already in the wild that only require a software update to reach FSD and AP can improve weekly, just like it has. Didn't see a video of auto Lane change, though might have missed it.
Edit: watching the videos again I notice all the cars passing the Volvo, how slow are they going. I have never seen such beautiful Lane markings in my life, certainly not that good where I drive. The one guy even says he has to help out around corners. I actually like that feature, auto resume. Not bad, but frankly Tesla created AP2 in 6 months as an answer to the fallout with mobileye, which the Volvo is using. That alone should give you confidence in what Tesla engineers can do. And we have not seen FSD1.0 yet.
Last edited:
According to those videos, Pilot Assist 2 handles curves better than AP2 definitely and probably even AP1. There are a lot of sharp turns in that video that its handles flawlessly. Secondly there have been not one single complaint of sudden breaking for signs, bridges and overpasses. Cars in front disappearing, truck lust, swerving lane lines, rear-ends and near rear ends. It also rarely takes exits and handles local roads okay. something AP2 struggles horrifically at.
As I have already pointed out, Volvo cars can be updated OTA with new software versions and map updates.
Volvo also already have cars in the wild, in the hands of 100 real customers.
An actual L4 car with hardware and software complete with no driver as backup that customers are using right now, today in sweden.
Meet the First Real Family Slated to Get Keys to Volvo’s Self-Driving XC90
Lastly ADAS today can be made in 15 days because of the advancement of technology, science and academics and has nothing to do with Tesla engineers. With the advent of deep learning you can create a L2 system in acouple weeks. Ask Geohut and Nvidia.
Its funny how Tesla leads when facts still says they are the worst at FSD and other competitors like Volvo and Waymo actually have finished and near finished software. Waymo has 500 taxis in phoenix with backup driver, Volvo has 100 cars in Sweden with no backup driver. GM is about to deploy 1000s of SDC and start their own taxi. But yet again Tesla is mysteriously ahead with software that disengages every 1 in 3 miles and also mysteriously ahead in L2 system and has "no competition".
Reciprocity
Active Member
No one has a clue what Tesla FSD is or how far along it is as there is only a 6 month old video from Tesla which does clearly show that it is far more advanced then any AP. And that is the hardware now in 70,000 cars. No one has that, period. Not 509 Waymo mini vans, not Volvo's 100 cars. No one. But I guess you are probably right and Elon is wrong when he says a demo of NY to LA fully autonomous is on schedule for Nov-Dec 2017.
JeffK
Well-Known Member
We do have a clue of what Tesla FSD is. Its called California DMV Disengagement Reports. I'm having a hard time believing how a demo that breaks down every mile better than what Waymo and Volvo actual have?
Waymo system that works 1 in 5,000 miles in suburban/urban environment and probably will be at 1 in 10,000 by end of the year (based on their improvement trend).
GM cruise that works 1 in 400 miles in dense crowded urban environment and probably will be at 1 in 5,000 by the end of the year (based on their improvement trend).
Volvo says their L4 highway system works 1 in 310,686 miles.
Autonomous Vehicle Disengagement Reports 2016
From what you saying. Waymo can remove their back up driver by the end of 2017. Launch full scale taxi in phoenix and california with 2,000 minivans and you will still be saying that Tesla is ahead. Isn't that illogical and bizarre?
Last edited:
None of them are L4 cars that doesn't require human driver nor allow drivers to sleep in them. Volvo is actually selling directly self driving cars that you can sleep in right now (no software update required) and you frown against it? isn't that amazing?
Thanks for the info. I'm glad that Volvo is adding some competition.
I purchased a Tesla because of the styling and the power train. The AP was a cool adder.
I did test Honda/Acura LKAS and it wasn't very good.
JeffK
Well-Known Member
Correction... The Drive Me program cars go to zero customers, 100 families will get to ride in them, and an engineer sits behind the wheel the entire time, like uber.
They mention it here
Volvo to seek volunteers for self-driving car trial in UK
Last edited:
Thanks for bringing up the California DMV Autonomous Vehicle Disengagement Reports 2016. I haven't had any time to keep up with this thread but from the above URL, those who are making arguments about Tesla vs. Waymo (err... Google errr Alphabet) and others should look at how many autonomous testing miles on CA public roads Waymo has done vs. Tesla in the 2016 report.
In 2016, Waymo did over 635K miles on CA public roads and has 0.2 disengagements per 1K mile w/124 reportable disengagements. For 2015, they did 424K miles w/341 reportable disengagements at a rate of 0.8 disengagements per 1K miles.
For 2016, if my math is right, Tesla did only *550* miles and did it only in October and November. I didn't count the # of disengagements, but my guesstimate is somewhere over 150 disengagements. For 2015 (see Autonomous Vehicle Disengagement Reports 2015), they reported no disengagements. I can only presume that meant they did 0 or near 0 miles of such testing on CA public roads.
Yeah I'd like to see a comparison in Southern California or other places with really poor lane markings. The head of Volvo lost his temper last year in front of a reporter and LA's mayor when pilot assist I kept losing lanes. "Why don't you paint your bloody roads?" was the quote if I recall.
Anyway competition is great and can only help push Tesla forward faster.
Similar threads
