Google veteran pans Tesla Autopilot: “We were doing better in 2010”
arstechnica.com
arstechnica.com
It doesn't matter how good your invention is if you keep it to yourself.
That, or one month before the big announcement someone decided to scrap it all and transfer all engineers to G+
You can throw on a lidar and radar on a car, along with a database of high res maps to get next to perfect 3d representation and path generation. That tech has been around for a while.
On the other hand, a camera based system is cheap, through the nature of it being specifically imperfect, you have to make systems that work with imperfect data, which make them more robust.
> Urmson predicted to Bloomberg that Aurora would be the first company to deploy self-driving technology for long-haul trucking routes at a "meaningful" commercial scale.
Tesla could do this tomorrow with Tesla Semi. Their hold-up is manufacturing and deploying a new electric vehicle, not developing self driving software which they already have. This is opposed to every other self driving startup in the space which has no practical driving experience and no commercial product to deploy on.
To say that meaningful advances in the space have not been made in the last decade is kind of weird. We are still far away from L5, but closer than we were before.
From my previous personal experience, once you start having discussions like these, the original poster gets hostile, moves the target post... etc.
Yes, it does just that (unless you buy the FSD scam), but it also feels like an integral part of the car instead of a addon for higher trim models. Everything (including the tablet) is centered around autopilot and it's natural to use anywhere. Would I use lane centering in a Honda or Hyundai? Maybe. I use the tempomat (cruise control) in other cars too, but I never felt comfortable in using just that to drive all the way to SF.
I don't think I would do that in a 2010 Kia. I might be wrong.
Tesla lessened my mental burden and I just watch other cars while it does its thing. I am also more confident as it watched other cars by itself using the cameras, it's like 10 pairs of eyes instead of 2.
Lane centering does not allow you you totally take you hands off the wheel and let the car drive itself. It’s just a small assist to help once you’ve almost already crossed into someone else’s lane. To be fair, autopilot is not supposed to be totally hands off, you’re supposed to still touch lightly touch the wheel every few seconds. You still need to monitor it.
Adaptive cruise control is pretty nice, but only useful on high speed highways.
They aren’t even comparable really. The Tesla is truly driving itself, while the other is just an assist in case you mess up. Also Tesla is beta testing the next version of autopilot that can drive in cities, and obey traffic signals.
But, we are rapidly approaching a time where lane assist & adaptive cruise are becoming ’standard’ on most vehicles and I think that’s great.
The lane keep assist has a wide band it allows the car to ride inside. At first I was disappointed in this, I wanted it to keep the car perfectly centered at all times. However it makes sense when you consider how the car detects that you're still alert and watching the road: you have to apply some torque to the steering wheel once in a while. If the car was perfectly centered already there would be no need to do that.
I think maybe a combination system that detects gaze and not only steering adjustments but even just touching the wheel without applying pressure would be better. It would allow to lane keep to do better while still ensuring driver attentiveness.
My girlfriend’s 2021 Outback has pretty amazing adaptive cruise, that will come to a complete stop as you describe. But it doesn’t work under a hard cut off of 35 miles an hour, and trying to use it outside of a highway situation is very nerve racking, which is why I said it’s not ‘useful’.
These technologies are great, and I’m excited to see the become more standard. But my point was to compare them to Tesla autopilot is apples and oranges. To think that they are the same, or they are designed for the same purposes is to not appreciate how darn good Autopilot really is.
When I drove mr brothers model 3 I couldn’t really tell a difference.
With their upgraded auto-pilot the car detects and automatically passes other drivers, takes exits, can completely handle complex interchanges by itself changing lanes to find exits while getting in the correct lane, slows while passing, gives trucks extra space on the highway, can handle complex construction zones without leaving it's often confusing lane, automatically exits the passing lane when not passing, etc.
Just the precision with which it can change lanes in dense traffic is so relieving. I hate changing lanes without it. I now spend my time keeping a close eye on idiot drivers while the car manages the lane change minutia - don't hit the car up front, make sure I have space, make sure no one is passing on the right, make sure no one is cutting off the space im trying to use from the 2nd from the right, look forward, look back, look forward again to check traffic, look back - never again.
I inherently trust the Tesla to drive better than me most of the time. I inherently trust the Tesla to drive better than 98% of other drivers.
it only took a couple of deaths to train their ai for construction sites.
> Just the precision with which it can change lanes in dense traffic is so relieving.
parked white trucks still a issue tho.
this whole thing reads like a marketing brochure, while conveniently ignoring the trail of deaths the tesla gung-ho attitude about releases brought.
So, accidents like you describe are not to be unexpected. People get complacent and think it really is a self driving car when it’s not. But let’s not discount the amazing level of sophistication Tesla has achieved in this system. It’s really impressive, and head and shoulders above anything else available.
That being said, I would wager that the high profile accidents you mention were due to drivers being negligent and abusing the system, where they should have been alert and paying attention like we all do when we drive.
You mentioned ‘bouncing off the lane edges’, and the reality is this is still where most companies are at. Which is fine because that’s all it’s supposed to be. It’s designed to be an assist when you mess up. Just a gentle nudge back to where you’re supposed to be.
I’m glad this stuff is becoming standard though it’s great tech. My point though is they are not meant to be ‘self diving’ is meant to be a safety ‘assist’.
Honestly, Elon still refuses to call it ‘self driving’ because the driver is still expected to be alert and monitor the autopilot. But the level of sophistication autopilot has is so far ahead of anything else, my point was it doesn’t make sense to compare them.
Pretty soon they will finish beta testing, and you will download an update, and the Tesla will be able to navigate complex traffic situations including cities and traffic signals.
To be fair, I saw something where Honda is supposedly developing a competitor to Tesla autopilot, but the current tech is not that. And the Honda autopilot will only be offered in Japan in a very limited quantity of some exotic $100K+ model. Who knows when those efforts will trickle down to be regular production vehicles available in America. But it’s available today in a Tesla. It’s really impressive what they have achieved!
In-fact most systems in the last few years all only "weigh" on the steering wheel, they don't actively steer the vehicle to return it to it's lane they just know "I know you don't want to go over that line without a signal" - very different and may not avoid an accident.
If you've actually put these systems to the test like Honda's system they've had since 2015....they suck. I test drove a fully loaded 2017 Honda Accord and the mistakes it made were atrocious and dangerous. I decided to just wait for a Tesla at that point.
Case in point: God forbid you cut 1ft of the yellow line at a highway exit or the Honda will yank your steering wheel so hard you move all the way beyond the white line on the other side of your lane due to the curvature of the actual road not being accounted for properly by the naive emergency road departure system.
If that is true, there must be numerous lawsuits and recalls to account for all the deaths.
I appreciate you being nice. I guess all Tesla posts are incredibly controversial so we just assume that posters fit into one of several predefined camps. No worries, I understand!
> 2010 on a Kia
So according to the Wiki article that would be the 2011 Kia Cadenza which I cannot for the life of me find the MSRP of. Please let me know if you do. What I could find is the MSRP of the 2021 version, which starts out at $37k (notably still more expensive than a Model 3). In addition, lane departure warning is different from lane keep (I'm sorry if this is splitting hairs but I think it is a big difference); with a warning you still have to take action but with lane keep the car will drive back to the center of the lane for you. This removes the driver's reaction time from the picture, which of course is one of the primary goals of lower-level self driving. Of course, you could argue that lane keep assist is standard on a lot of new cars today, which is my point; it wasn't standard back then, it is now, likely at least a little of that movement is due to external pressure from Tesla.
> It is hard to assess your argument that Tesla can deploy their semis tomorrow... etc.
I can try to expand on my argument:
While a semi is obviously a much different vehicle type than a sedan, Tesla already has experience deploying to different types of vehicles (S being a hatchback, 3 being a sedan, X being an SUV, Y being a crossover etc). Of course a semi is heavier but the way it drives is still the same, it just has to brake earlier than most other cars (just like an SUV at the same speed has to brake earlier than a sedan due to the weight difference). I don't claim to be an expert on self driving vehicles, but it is my perception that the physics of the car don't matter as much as it matters for the model to be able to detect and respond to stimuli, which can then be translated to physical movements for electric and hydraulic linkages. Therefore, if a Model S can detect an obstacle in front of it today and brake to avoid hitting it 99% of the time, I don't see why that same system can't be applied to a truck that weighs differently and have it still work. Yes, there is now a trailer in the mix, but Tesla owners already use Autopilot with trailers attached (for better or worse - I don't know if that's recommended or not).
Now, we could split hairs over what Urmson meant by "meaningful," but for me, meaningful self driving long haul driving translates into self driving on the highway. This means that a trucker could manually drive from depot to highway, switch on Autopilot, and then monitor the road for the next few hours until their shift is over. This is obviously not as good as being able to "set it and forget it" - such that the driver could go to bed or something like that - but I would argue it is still meaningful and upends the status quo for commercial drivers today.
Could Semi be deployed tomorrow? No, obviously, it doesn't exist outside of a few demo models that may or may not have Autopilot support yet. But my main argument is that it's being held back by the "car" stuff that we already know how to do (we know how to build electric trucks), and not the self driving stuff that is much more up in the air.
Frankly I don't know if Semi will be a good self driving vehicle - we of course have to wait and see - but what I do know now is that it will be the only self driving truck to have all of the features that Autopilot offers when it comes out in the next few years. I don't think traditional truck manufacturers are going to be able to catch up in time, and I chalk that up to Tesla having the capacity to gather experience on the roads when no one else can (see my other post: https://news.ycombinator.com/item?id=26839891)
I hope this helped explained my thought process more. For what it's worth, I do have a positive outlook on Tesla, I do think they make the best car for a person like me (young software developer, not really materialistic), and I hope to be able to afford one some day. But, I don't think it has to stay that way. I feel this way about Tesla specifically because I'm just not seeing other automakers invest in the areas I care about: green energy, infotainment and automation. This is probably because I drive a gass guzzler SUV with no center console and no computerized safety features except for ABS and traction control. Other people who value other things will of course like other cars.
I feel that this refutes you point "This is opposed to every other self driving startup in the space which has no practical driving experience and no commercial product to deploy on."
However, Waymo is using outfitted consumer minivans, not custom built vehicles. You can imagine that if the system becomes popular enough Honda and Toyota are going to want their slice of the pie once they realize that they can rent out self driving models rather than selling once at a dealership. It will be easier for a car maker to work with existing software rather than for a software company to build a vehicle. Tesla here owns the full vertical stack, much like Apple does with smartphones. That doesn't mean that they will control the entire market, however they will be able to squeeze much more value out of the market than competitors.
https://arstechnica.com/cars/2020/10/waymo-finally-launches-...
I’m not sure this is boost you really think. If I test you on your times tables every day that doesn’t teach you calculus. In the same way I think the vast majority of data Tesla is receiving is reinforcement of what already works.
> the vast majority of data Tesla is receiving is reinforcement of what already works.
Absolutely! But is this not still valuable? This tells the model that, in times when it doesn't know what to do, it can pick the most reliable strategies that have worked 100% every other time. This applies geo-spatially too: if Autopilot drives too hard around a corner and you hit the brakes to disconnect, the model will remember this and send the incident to the cloud. As more Teslas continue to disconnect around this corner, the model will learn to adjust its speed for a smoother ride until nobody is disconnecting. At that point you, someone who has never driven a Tesla before, can go out and buy one and it will handle that same corner perfectly without you knowing that it was once a pain point. This is the benefit of collecting as much data as possible.