I don't think any one is doing this. This is the point Karpathy is making. Vision is essential and needs to be perfect and also has all the information to recreate the 3d environment. That's why lidar is not required and is expensive.
lidar is not required and is expensive.
[1] 2019 https://arstechnica.com/cars/2019/03/dashcam-video-shows-tes... [2] 2018 https://news.ycombinator.com/item?id=16772748 [3] 2017 https://bgr.com/2017/03/02/tesla-crash-video-texas/ [4] 2016 https://www.thesun.co.uk/news/1402992/tesla-self-driving-car...
I didn't say anything about humans vs self-driving cars. I'm arguing about self-driving cars with LIDAR vs those without.
The status quo is humans so the first question is whether it can do at least as well as humans even without LIDAR. It seems to be doing that, even if the things it does get wrong are different things than the things the humans get wrong.
The question of LIDAR vs. not is a separate issue which is purely cost/benefit. Even if LIDAR has some value, it remains to be seen whether that value is ultimately more than its costs, in the same way that we know how to make cars safer by making them weigh 8000 pounds and yet in general we don't do that because the cost is more than the benefit.
The public road right before the drive to my house has an s-turn in it. The CWS goes off about half the time I go through that series of turns just hundreds of feet from my turn. I can see the turn in the road and the guard rail ahead, the car sees a stationary object I'm driving straight towards at around 45mph. If it automatically braked with its current awareness, I might never get home or get rear-ended by people following.
We are trying to build a car that doesn't crash.
If you want a car that doesn't crash, never build the car let alone drive it. Then it'll never crash.
The point is both reduction in fatalities and increase in human quality of life. Not an impossible perfection.
Why is it one or the other? Why only a human driver or a robot driver? Could you have a human driver in charge with an accident preventing robot, getting the best of both worlds? If the human falls asleep, the robot can do trivial driving, and the human still makes the more difficult driving decisions.
The question isn't if robot car is better than an average car being driven by a human. It's if the same robot car would be safer if the human was driving it, and the robot in accident prevention mode. And I don't think that is going to be true in any near future.
We already have this and we've had it for years. What do you think automatic emergency braking is? Adaptive cruise control? Lane keeping assist?
> The question isn't if robot car is better than an average car being driven by a human. It's if the same robot car would be safer if the human was driving it, and the robot in accident prevention mode. And I don't think that is going to be true in any near future.
No clue what you're trying to say here. Driver assist features don't improve safety?
I am tired of seeing driver assists being sold as driverless systems.
Autonomous cars are one of those things that really need to be done right or not at all. The notion that human drivers will be ready and able to take control at a moment's notice is nothing short of laughable.
IMO, about the only way partial self-driving cars can work would be to delegate it to a remote building full of 'drone drivers' somewhere. Heck, I'd pay for that today, just out of laziness.
A lot of the utility of self driving cars is not just increased safety. It's allowing disabled people to be more independent, reducing the need for car ownership with (less expensive) self-driving taxis, reducing the need for parking lots and allowing that land to be used differently, etc. Do you think self driving cars should be prohibited as long as they are less safe than the standard set by human+collision avoidance cars?
I don't think that makes sense unless cars with human drivers and no collision avoidance are also prohibited. Even then I'm not sure. As a society we know that driving is not totally safe but allow people to drive anyway because it's a useful tool. I don't see why the safety standard should be even higher for self-driving cars, which are even more useful, especially for people who can't drive.
One last thing is that anything that slows the adoption of self-driving cars will probably lead to a much greater number of old, less-safe, cars in use for a longer period of time, since it's very unlikely politically that the existing car stock will be just kicked off the road, regardless of what regulations are set for new ones.
If having a human in the car reduces the accident rate, then I propose making it necessary for the human to be in the car. Are you advocating that we make roads unsafe so that disabled people be able to take taxis more?
> I don't see why the safety standard should be even higher for self-driving cars.
I see it as the opposite. The advocates for driverless cars are advocating lowering standards. For me the standard is simple. If the same car is as safe driving itself as it is assisting a human driving it, then it can drive itself. You can only remove the human if the human adds no safety in the same car. Please do not compare to "the average car on the road".
I see there are positive externalities to cheaper taxis. There are really negative externalities too, in terms of traffic, urban sprawl, and carbon emissions. But I don't think we should deviate from the topic of safety for those externalities
I disagree with those, so I also disagree with the premise that having a human in the car should be required for any amount of increased safety. I think the amount matters. And keep in mind we are only talking about safety increases from present day, you're just proposing to ban them in order to achieve a greater safety increase.
My >90 year old grandmother still drives. She hasn't run anyone over, but I imagine her reaction time isn't the best. Apparently it's common for cops to be lenient with elderly folks who are known bad drivers, since in many areas they wouldn't be able to get groceries or go to the doctor otherwise. I'd argue that maintaining the status quo (by having a blanket policy) for many groups of drivers would probably be safety theater instead of anything real.
Lastly, I touched on this in my earlier post, but if you care about deaths it's very important to consider what cars are actually being replaced in addition to individual car statistics. Imagine in scenario A, 10 million older cars are replaced by 10 million driver-assist cars, and scenario B where 30 million older cars are replaced by 15 million self-driving taxis. Assume the driver-assist car is safer than the self-driving car, which is safer than the older car. Even so, the driver assist scenario may have thousands more deaths, since the uptake rate is different. And this is without taking into account any safety-improving software updates, which are very difficult to estimate but may be very significant statistically.
A simple per-car comparison may lead to all of the negative externalities and thousands of more people killed. I really do not see how that outcome could be justified in the name of "safety". At the very least, even if safety is the first and only concern, policy makers should try to minimize vehicular deaths which requires considering things at the fleet level.
Basically, Musk's point is that if self-driving cars are statistically safer than humans, then allowing humans to drive will make cars less safe again.
Furthermore (regarding some other children here), the study that Lex and his team has done show that drivers are not less engaged while Autopilot is on (as many critics say).
You mean "if driverless cars are statistically safer than humans". It shouldn't be a surprise that driver assistance decrease accidents. But Autopilot isn't a driverless system, and neither is there satisfactory proof that it is safer than a human.
I'm not sure that human+robot would be better than robot alone. Eg. The human is in control and starts moving close to the adjacent lane without indicating. Does the robot (a) assume the human driver is changing lanes and so turns on the indicator light; (b) take control of the car and move it back to the center of the current lane; (c) do nothing and let the driver change lanes without indicating? The ambiguity in the human driver's intentions means the robot might make a decision that is less safe than if it was in complete control.
Lidar gives you very rudimentary depth and positional data, but it is precise, and reliable, and the false positives are low. Lidar is not seeing the world as a series of probabilistic phantasms that may or may not be a pedestrian, or a jersey barrier, or whatever.
Elon said: "Lidar is lame! Lidar is Lame! lame lame lame lame."
But then he later said: "We've gone over this multiple times. Are we sure we have the right sensor suite? Are we sure we have everything we need?.... No."
Lidar is a crutch and we'll see how hard Tesla is limping in a couple years when Elon is claiming They're going to have Robotaxis.
To Tesla's credit, it is theoretically possible to develop a fully autonomous vehicle without needing Lidar, but before you can deploy, you need to integrate Lidar, and it can be retrofitted as a largely independent system, necessary for it's enhanced redundancy, but not needed for informing the planner, except when the Lidar is seeing things the vision system isn't.
The presentation was a pretty good run down on various techniques and methods that are being utilized by different autonomy developers. The downside is that everything is framed as a rationalization as to why Tesla has some kind of secret weapon, when they're mostly doing conventional things, and not all the conventional things you need to build a complete autonomous vehicle.
I was waiting for Elon or Karpathy to trot out the magic pixie dust that will have Tesla's driving empty in a couple years, but that was nowhere to be found.
Waymo's safetly critical disengagements are at 1 every 11,000 miles. Fender benders happen about once every 100,000 miles, so just on that Waymo is off by an order of magnitude, and Waymo spares no expense on compute or sensors.
Is there much of anything a car can do to prevent being T-Boned?
And even if there is, I don't really see it being something that radar would help much with. If you and another car are approaching perpendicularly, you'll have an almost perfect situation to get positioning and speed information over a pretty short amount of time.
Of course I've never done anything even resembling any kind of work in this space, and I don't want to pretend to know the answer, but the claims of side radar not being necessary seems to pass a smoke test.
Yes. Stop suddenly while entering the intersection when conflicting traffic is detected. Google cars do that, and they have been rear-ended several times because of it. You can find those events in the CA DMV accident reports. Better than being hit by the cross traffic.
Not that Tesla is doing any of that right now, they most likely are not. But a lidar is not going to, either. Not ever, which is the point.
I thought the fleet learning aspects Karpathy outlines came pretty damn close to the pixie dust.
With regards to fleet learning, Tesla has been going on about it for years, but more data is not better. Better data is better. Clifornia's disengagement reports have seen criticism because they place an emphasis on the number of overall miles driven as a useful measure of progress, when actually it doesn't mean much. Companies like Zoox, after spending years doing small-scale closed course testing hit the ground running, and with only 10 vehicles and a few months of driving around San Fransisco they were able to take reporters for autonomous rides, and they took a couple dozen journalists to various pre-designated spots around SF with only one disengagement the whole day.
The nice thing about the whole fleet learning narrative is that is makes Tesla drivers feel like they're helping just by driving around.
EM radiation from UV down to radio has different propagation characteristics, and materials form different attenuators. Radio easily passes through most non-metals with less attenuation than 400-800 THz radiation (light). When I was at Tesla, I tried like heck to get people to internalize this, but the 'cameras + one forward-looking radar is enough' crowd has clearly won.
Heck, RF can shoot through mud, rain, sleet, and snow. Cameras? not so much. Although I agree that the number of 'RF pixels' one gets is nowhere near what a camera provides, it does provide some pixels when the cameras are covered in road muck.
And what about when it's dark? Are the headlights going to provide enough light for the cameras to function well in all road conditions?
What an amazing position to have been in history, even if it didn't go according to the way you felt (and imo probably are) was right.
I feel like this would be able to detect trees in the distance and other objects to orient the road/lane detection. And keep it cheap. Possibly add on to existing cars.
https://hn.algolia.com/?query=solid%20state%20lidar&sort=byP...
Where did he say this? He repeatedly said they have everything they need to drive.
>but before you can deploy, you need to integrate Lidar,
Why?
>not all the conventional things you need to build a complete autonomous vehicle.
No one knows how to make a completely autonomous vehicle because they don't exist. I'll take the billionaire tech genius' approach, whose gotten farther than anyone in this field, over your armchair opinion, thanks.
What they've done so far is a dangerous lane-assistance. From what's seen they are nowhere near waymo. Would hate for tesla and uber to ruin this for everyone else.
Lidar is inherently probabilistic too,like every sensor is. Just because the photomultiplier voltage rises and some peak detection hardware can correlate it with a transmit pulse, it does not mean that there is a "point" at that position in 3d space. It has different error characteristics to vision sure, but it is still an inference problem with corresponding artifacts to deal with. In fact I would say that the different error characteristics actually make them ideal for sensor fusion, i.e. Lidar for initial depth and vision for filling in the details.
As far as I can tell, LIDAR makes a lot of prediction tasks much easier (since you have both images and LIDAR point clouds), and is also a backup for obstacle detection to ensure that your self-driving car doesn't crash.
It also doesn’t have all the information to recreate the 3d environment. There’s all kinds of inferences that human drivers make that require a good model of the world, to start with. For example, a car shouldn’t brake for a small object that moves with the wind, even if it looks like it’s made of concrete, or looks like a child.
And not sure who you are but I trust all of the other companies who are doing this a lot more. And they are all using LIDAR.
It's a mistake to think LIDAR doesn't have these drawbacks - its an optical system at its core.