Slashdot Mirror
GM Exec Says Elon Musk's Self-Driving Car Claims Are 'Full of Crap' (smh.com.au)
An anonymous reader quotes the Sydney Morning Herald:
Billionaire entrepreneur Elon Musk's claims about the self-driving capabilities of his upcoming Tesla vehicles are "full of crap", General Motors' self-driving Tsar says... "To think you can see everything you need for a level five autonomous car [full self-driving] with cameras and radar, I don't know how you do that"... GM's own solution involves several radar and Lidar sensors, as well as cameras and multiple redundancy systems. Each system costs hundreds of thousands of dollars, and GM are some way away from getting the cost low enough to be commercially viable. "The level of technology and knowing what it takes to do the mission, to say you can be a full level five with just cameras and radars is not physically possible," Mr Miller said.
I definitely trust GM to make an unbiased analysis of competitor technological capacity.
you're funny, the list of things computers can't do with any amount of sensors, that humans can, is quite long
I'm guessing Tesla really can't do it well enough, cheap enough either ... yet.
But one of the advantages of having a Bond villain as chairman and CEO is that he's a little less bound by quarterly profit targets and the need to dole out healthy shareholder dividends like clockwork.
For the first fifteen years after Microsoft went public it never paid a penny in dividends. Investors didn't expect dividends; they expected all the profits to be plowed back into world domination.
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
But to clarify the difference:
LIDAR: Formerly about $75k, now about $7,5k per unit, and requires a bubble dome on top of the vehicle. GM and Waymo use it, Tesla doesn't. In addition to looking weird and adding drag, the price is killer if you want to include something on every vehicle. Beyond this, LIDAR doesn't work in fog, heavy rain, snow, and other conditions that humans can drive in - meaning that you'd have to either prevent trips during these conditions, require humans to drive during them, find workarounds (not easy), or rely instead on other sensors. And you still need to understand the world around you visually - LIDAR will tell you that "something" is there, but it can't read signs, see road markings, see brake lights, tell if that thing in the road is a person or a paper bag, etc.
Tesla, for these reasons, ruled out LIDAR. They just simply use the "other sensors" - 1x radar, many cameras, many ultrasonic sensors - all the time. This way, all of their sensors can be put in all of their vehicles, and do double duty for both self driving and standard safety features (autobraking, etc), depending on what options the buyer has paid for. This however comes at a penality: when LIDAR works, it works really well. Photogrammetry with cameras is prone to stitching errors, and radar, while being able to see some things that humans can't, sees the world in very strange ways (for example, a piece of plywood is transparent, but an aluminum can glows like it's on fire). It's a much more challenging task if you leave LIDAR out of the loop. But, it gives you a more saleable product.
In the end, I expect a convergence to take hold. An interesting new technology for example is time-of-flight cameras - they function as normal cameras, but also can read the length of time it takes for a laser pulse to return on every pixel they record. So no dome, just your normal camera coverage and a few cheap, fixed lasers - in mass production, it might not cost much more than cameras alone. In such a case, I'd expect the LIDAR groups to simply replace their conventional LIDAR datastream with the time-of-flight datastreams, while I'd expect the non-LIDAR groups to replace their photogrammetry-and-radar built 3d models with time-of-flight 3d models. But both sides will still need image processing, so it's important to work on maturing that technology today.
That said, let me reiterate that I'm a pessimist regardless of what tech you use. There's just so much nuance in driving in hazardous conditions - understanding when, where and how much you have to slow down, what's safe to drive on and what's not, what things to the side of the road are hazardous and what aren't, when you should break rules (such as driving in the middle of the road when conditions are dangerous but oncoming traffic is rare), what are the consequences of a mistake in one location vs. another, etc, etc. On my gravel road, there's a canyon to one side with no guardrail, and varying amounts of ice and potholes in different places. You better well know how your traction is going to fare as you move across the potholes (vibrating the car and making it lose traction) or icing if you don't want to end up in an unrecoverable slide into a ravine.
Just to pick a random example among countless things that you have to take into account: how long do you think before any self-driving systems will have "sheep recognizing algorithms"? Because where I live, there's sheep. Group of sheep on one side of the road: probably safe. Group of sheep on both sides of the road: not as safe, but probably safe. Lamb on one side, ewe on the other? Very dangerous - the lamb will invariably run to its mother as you approach. Where's the ewe-lamb-running algorithm?
"If there was an antonym to 'Elon Musk', it would be 'Richard Branson'."
Also, as to answer the question of how humans do it with "two cameras": logic. We don't have "stitching errors" in how we build up a model of the world around us from visual data because our brain constantly processes everything around us through the prism of "does that make sense?" But whether something "makes sense" or not is an AI-hard problem.
Building up 3d models with photogrammetry is an inherently error-prone process because a computer doesn't know if something makes sense. The approach is "Oh hey, these patterns from the left and right camera matched up - there's an object there at X distance based on how far the patterns had to be shifted to align". But what if the patterns happen to be different things that just happened to match up in patterning? Or what if, due to lighting / texturing / obstruction / material issues, the same thing didn't look exactly the same from different angles? Uncovering these problems is, as mentioned, AI-hard. I doubt anyone is even trying at this point; there's enough to work on just to get things to follow road lines correctly and not go chasing old tire tracks or poorly erased construction lines.
Ranging systems like LIDAR help let you just simply ignore the issue by telling you flat out, "I sent out a beam in this precise direction and got a reflection in precisely this length of time." But LIDAR has a number of problems, as described above.
"If there was an antonym to 'Elon Musk', it would be 'Richard Branson'."
According to Google (who's a big LIDAR proponent), it's still $7,5k per unit. It still messes up your aerodynamics and looks dorky. It still can't see in adverse weather conditions, meaning you have to have developed an optical / radar based world-modeling system anyway. And you have to have image processing regardless to read signs, road lines, identify objects, see brake lights, and so forth.
There's real hope for further improvements in LIDAR and its variants in the future, however. We'll see where it goes.
"If there was an antonym to 'Elon Musk', it would be 'Richard Branson'."
Also, we crash more often than is permissible to a self-driving vehicle.
Microsoft demonstrates the transition from a startup, growth oriented company to a mature, profit oriented one.
Jobs, by the way, hardly counts as an inventor. Visionary, sure, but that's not the same thing at all.
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
Next you're gonna be telling me how we're not all going to have villas on Mars by the end of the next decade, at the very latest, all brought to you by SpaceX brand rocket ships.
Elon gets my respect for making two successful and innovating businesses that have lasted and have solid fundamentals into the future. He gets no credit for his bullshit factory. Good bullshit has to be believable. Self-driving cars and 800 mph trains in a tube by next year doesn't pass the giggle test.
No, that is not what he is saying. What he is saying is that GM cannot do it with these limitations and there is very good reason to believe that others cannot do it either. As Musk is full of it in a number of topics, it would not surprise me one bit if he were on this too.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
A person gets all the information about where he car is and has to go based on two sensors (basically two cameras with stereoscopic vision) in your head, positioned sub-optimally inside the vehicle, with one point of view at any one time, and sensors for speed. If you include the persons ass, throw in a cheap accelerometer too.
There is no reason to think cameras and an accelerometer canâ(TM)t figure it out with software to the same degree. But the cars cameras have better vantage points, near perfect operation once the software comes around, and will emphatically understand the rules that govern the roadways better than we could, as well as the dynamics and limitations of the vehicle it is operating. Cars can absolutely get autonomous with less than GM claims.
The human brain has a huge amount of computational power compared with the processors in a car. And typically you're not allowed to drive a car until you're a teenager. So, that brain has 15 or more years of training in identifying all those objects you see while driving. Chances are good that the brain has been sitting in a car many times over those years and gotten good at identifying them at speed.
It's definitely possible that we'll eventually get sensors that can do that, but it's naive to suggest that we're anywhere near that point. The sensor arrays can easily miss a bicycle or motorcycle if it's positioned in the wrong part of the road. With sensors that actually cover the entire lane ahead far enough to cover the stopping distance, it wouldn't be much of an issue, but most vehicles have far too few beams for that to happen. They have gaping holes right ahead of the vehicle that wouldn't exist for a driver. Drivers mostly can't see the couple feet ahead and behind, which are only an issue when going slowly. At speed, you wouldn't be able to stop quickly enough to care about that.
It's not a simple matter of can or can't do. The problem is there's no standard threshold of success which needs to be met for a system to be considered a "marketable" autonomous car. If your car can handle 95% of situations, is it suitable for use on the road and for sale to the public? 99%? 99.999%? Or maybe the proper metric isn't situations, maybe it should be average time in operation before it encounters a situation which stumps it. Should that standard be 1000 hours (6 weeks)? 10,000 hours (a bit over a year)? A million hours (over 100 years)?
Without some sort of standard, you can put a brick on the accelerator and a bungee cord on the steering wheel, and call it an autonomous car. Because it is, for about 20 seconds before it drifts into the next lane. It sounds like GM is working to a much more stringent internal standard for autonomy than Tesla, and the GM exec is frustrated that the press is constantly comparing them as if they were equals. Whether or not the car can drive autonomously isn't as important nor relevant as how often it fails to drive autonomously.
All you people who love government regulation should be all over this, instead of giving Tesla a free pass just because they're Tesla. It's why we have nutrition labels, Energy Star labels, NHTSA crash safety tests, EPA mileage ratings, standardized health plans under the ACA, etc. So buyers can easily compare products on a like-for-like basis
I'm sorry, but as an ex-roboticist, I have to disagree. The centralized approach is actually expected to be significantly more difficult.
Driving isn't the problem. There have been partially-automated train systems for decades, and some fully-automated ones more recently. We can easily make a system, centralized or not, to get travelers to their destination. The far more difficult aspect is dealing with the unpredictable interference. Children run into the streets. Animals think train tracks are comfortable beds. Storms knock down trees and flood routes. Every winter, potholes turn small cracks into large hazards.
To detect those problems, we have two approaches. The centralized approach is to have a vast array of sensors constantly monitoring every foot of roadway. That's a lot of sensors, so even being cheap inductive loops still puts the total cost in the billions. Unfortunately, "cheap" and "secure" are often mutually-exclusive. If the sensors can be hacked at scale, reality starts permitting the movie plots involving forcing an armored car or emergency responders to take the criminals' chosen route. Of course, with that many sensors, you also need a massive infrastructure project (and budget) to handle the input. That centralized coordinating computer has to be a supercomputer, even with modern processing, just to properly handle the ever-changing status of the roads. That's not even including the routing and coordination aspect, which would also need to scale as people are traveling. Coordinating a few million vehicles in a hurricane evacuation is no small feat.
Fortunately, the other approach makes a lot of those scaling issues disappear. By having a swarm of autonomous vehicles, the total sensing domain is limited to what's in the vehicle's immediate area, and to a lesser extent what will be in its future route. Rather than monitoring the whole road space, each vehicle can monitor just the road it's interested in. Detected hazards can be communicated to other vehicles, but that's merely advice given out of courtesy. Each vehicle looks out for itself, and as such the available processing capability naturally scales with the processing capability that is required. As technology improves, the new technology is deployed with new vehicles, remaining compatible with old vehicles operating in the same shared roadway, including those with no autonomous function. They are treated as minor hazards, just like any other object in sight.
You do not have a moral or legal right to do absolutely anything you want.
If you could somehow drop a human driver into the driver's seat in the middle of a journey with no prior knowledge, you would find them mostly non-functional. Even if they avoid panic, they would take seconds to minutes to bootstrap their knowledge to the point where they could effectively take over. The closest practical example I can think of to this scenario is how pilots are trained to deal with visual and inertial disorientation, and how they have to essentially troubleshoot their way back to an understanding of their situation. This doesn't happen in milliseconds, and in fact can take more time than is available as we've seen from well publicized air disasters. One can argue that the high rate of accidents from drunk and distracted drivers is due to similar disruption of situational awareness and an inability to recover with just two eyes and an ass in the seat.
A human driver relies on their mind much more than their sensors. Their mind builds a very elaborate world model based on sensory data over time and their understanding of where they are, what they have been doing, and what they expect to be the rules of the environment. The situational awareness that we "see" as a 3D world lit up around us is mostly a construct of our minds showing us our memory and our expectations. This includes not only some physical simulation and prediction (e.g. instincts about momentum and continuity of trajectories of objects in the scene) but psychological and social simulation (e.g. assuming intentions of other drivers and pedestrians and reading "body language" cues). Today's infatuation with neural nets and "deep learning" does nothing to tell us how to construct a synthetic mind with these sorts of abilities which are necessary to compensate for our paltry sensors.
This is the fundamental disconnect. A traditional engineer will think about how much data he needs from a sensor suite to reliably assess the scene with live data combined with the very primitive state model he knows exists in his automation system. A true believer in near-term AI will wave away the vast gulf between current technology and the human mind which we all take for granted every day, assuming that somehow the system can perform as well as us (or better!). More concerning, we have no reason to assume that a futuristic automated system made complex enough to emulate these functions of the human mind won't also be subject to analogous failure modes like confusion, delusion, hallucination, and even antipathy.
Elon's had a long history of proving naysayers like this wrong. My money's (literally) on him to pull this off. The folks truly terrified of self-driving cars are the National Auto Dealers Association (NADA) who stand to lose the most when you can order a car online and have it deliver itself.
No. there is an addtional pair of sensors that are critical- especially to accident avoidance.
The ears.
Nope. Deaf people can drive, and they have no higher rate of accidents than non-deaf people.
The issue is that computer vision doesn't work the same way as human vision. Human's are good at recognizing when things don't make sense, or spotting objects that are partially obscured and recognizing what they are. Humans know that when they can't see most of that thing because of the blinding sunlight reflecting off it, it's a car. The human eye has really good dynamic range too, and a built in self-cleaning system.
const int one = 65536; (Silvermoon, Texture.cs)
SJW, n: "Someone I don't like, and by the way I'm a fuckwit" - AC
Obviously it's the guy who drives on the median
According to Google (who's a big LIDAR proponent), it's still $7,5k per unit. It still messes up your aerodynamics and looks dorky. It still can't see in adverse weather conditions, meaning you have to have developed an optical / radar based world-modeling system anyway. And you have to have image processing regardless to read signs, road lines, identify objects, see brake lights, and so forth.
There's real hope for further improvements in LIDAR and its variants in the future, however. We'll see where it goes.
Good lidar systems see much, much better than camera based systems in adverse weather. I work on FMCW lidar systems and I recall driving to work one day where the fog was so bad that I couldn't even find the road to my office. Once I got work, I turned on the lidar system I was working on and it imaged a building 100 meters away without issue. Road lines are trivial to identify in a lidar system since they have much different reflectivity than the road surface. Objects are also easier to identify because you aren't trying to pull three dimensional information out of two dimensional images. On an FMCW lidar system, you also get doppler information for free. You don't have to try and decide if an object is moving towards or away from you by comparing subsequent images. Every single point in the point cloud includes a meters/second doppler value.
I have to assume your familiarity is with those awful spinning Velodyne systems. They are utter garbage. No self driving car company that I've interacted with is even vaguely entertaining the idea of using them in a production car. They don't even really like using them in their mule cars but, until very recently, they were the only real option available.
The real problem with lidar is that people aren't good at consuming lidar data yet. Once they start to get some experience with it, I have zero doubt that lidar will be the primary sensor on the car. It's the only way you can really build a model of your surroundings with high accuracy, high refresh rate and high tolerance to ambient conditions. So, I actually agree with the GM guy here: Tesla is full of shit. They aren't going to make a level 5 autonomous car with cameras and radar.