Still no LIDAR? LIDAR was the enabling technology in the DARPA challenges, and it remains so today. I can think of one Tesla customer who would still have a head if his "self driving" car was equipped with LIDAR.
>Perception is a game of statistics. We believe it will ultimately be entirely possible to build a self-driving car that can get by on, for instance, cameras alone. However, getting autonomy out safely, quickly, and broadly means driving down errors as quickly as possible. Crudely speaking, if we have three independent modalities with epsilon miss-detection-rates and we combine them we can achieve an epsilon³ rate in perception. In practice, relatively orthogonal failure modes won’t achieve that level of benefit, however, an error every million miles can get boosted to an error every billion miles. It is extremely difficult to achieve this level of accuracy with a single modality alone.
>Different sensor modalities have different strengths and weaknesses; thus, incorporating multiple modalities drives orders of magnitude improvements in the reliability of the system. Cameras suffer from difficulty in low-light and high dynamic range scenarios; radars suffer from limited resolution and artifacts due to multi-path and doppler ambiguity; lidars “see” obscurants.
https://medium.com/aurora-blog/auroras-approach-to-developme...
Not even stereo cameras. That's surprising. With all that compute power, depth from stereo should work well. A 3 camera system with all cameras pointed in roughly the same direction and arranged in a triangle, not a line, is resistant to many illusions that will trouble a 2-camera system. You don't need 3 cameras most of the time, but when you need them, you need them.
https://www.teslarati.com/tesla-autopilot-software-v9-blind-... (Tesla Version 9 real-world blind spot test shows Autopilot’s 360° cameras in action)
Youtube video of 3D Autopilot perspective: https://www.youtube.com/watch?v=wypE4fC56bg
Of note:
> Particularly noticeable in Erik’s demo, though, was what seemed to be a slight lag in how other vehicles’ avatars are displayed on the Model S’ instrument cluster, particularly when they are overtaking the electric car. That said, considering that blind spot monitoring is utilizing video feeds from the side and rear cameras, these slight lags could be due to the system shifting from one camera to another. This was particularly notable at around the 1:25 mark in the owner-enthusiast’s video.
Which leads me to believe that the sensors (cameras, front radar, ultrasonics) are sufficient, but why the NN hardware upgrade is going to be necessary. Tesla bet that it was going to be cheaper to drive the cost of the processing hardware down vs being beholden to LIDAR, and it looks like they might be right.
That's what's so striking about this. The new hardware
can run neural nets really fast. So they'll be able to
do something that can do the job most of the time, but
will still screw up badly some of the time.
I think that Tesla is moving too fast for public perception. A person is smart, but people are dumb, panicky animals, and expected value and engineering [1] set speed limits and goals that are far faster than PR, c.f. Greenpeace's anti-nuclear stance being significantly responsible for global climate change. But "they have to never make any mistakes" is a crippling double standard.
Tesla is moving too fast for the quality of its own technology.
On the flip side every day humans make mistakes that autopilot would avoid and would be “obvious” to an AP system.
Should we judge the computers vs perfection? Judge them like we judge humans? If a system came out that could be retrofit on every car in the country and make 1% fewer mistakes than humans, but the mistakes it makes are “obvious” and stupid, is it an improvement? Would we accept it? What about 10% or 50%?
But this isn't the case. This isn't the case of a company launching a system that performs better than humans through extensive testing and carefully weighting the risks. This is the case of a company intentionally false advertising something and resulting in preventable human loss.
That's less than 1/3 the rate of average.*
I understand the comparison is not direct since autopilot driving happens on the highway and those are general miles, but it is the best number we have for comparison's sake right now.
These are traffic deaths per 100 million vehicle miles by functional system type. There is not a single road type on this list, including the lowest (urban freeways), with lower fatalities per 100 million vehicle miles than Tesla Autopilot (which is 1/3 fatalities per 100 million VMT).
You'd need to compare the rate of accidents of a Tesla with and without the system engaged, or at least the Tesla compared to other luxury cars.
Which wouldn't even be that relevant, because autopilot features have been available for much longer in other cars.
And that's before you include the fact that autopilot hands over the control to humans, so, again, the numbers aren't comparable.
I understand your desperate need to validate your internal assumptions, but you don't have external data to do so.
[1] https://www.autonews.com/article/20170604/FINANCE_AND_INSURA...
> On the flip side every day humans make mistakes that autopilot would avoid and would be “obvious” to an AP system.
> Should we judge the computers vs perfection? Judge them like we judge humans? If a system came out that could be retrofit on every car in the country and make 1% fewer mistakes than humans, but the mistakes it makes are “obvious” and stupid, is it an improvement? Would we accept it? What about 10% or 50%?
When we can meaningfully charge a computer with a crime (or better, the developer of the software that caused the crash!), maybe then will public sentiment be aligned with the reality you envision.
If you want perfect, we will never get autonomous driving and 35k people a year die in the USA.
What if there's 10 Tesla fatalities a year, but similar people driving similar cars would have expected 50 fatalities?
Granted Tesla hasn't proven this, but generally there's going to be more deaths and we shouldn't just kill off the technology because they aren't perfect.
Keep in mind the decapitation was from someone watching a Disney DVD instead of using the AP like it was intended.
Ok, so you save a few bucks, kill a few people... then eventually come to the same conclusion every other Autonomous Driving company already has... you need LIDAR for this work reliably?
Humans don't rely on any sort of LIDAR to drive, I don't see how you could be so certain that computers need to.
We're nowhere close to emulating what goes on in a human brain... neural networks are an absurdly simplified approximation of simple life-forms... it took the K supercomputer (having 705,024 cores and 1.4 Petabytes of RAM) 40 minutes to emulate 1 second of 1 percent of human brain activity.[1]
Since we cannot come close to sensory input processing a human can do... we need to augment our systems with better sensors. LIDAR is a better sensor, and can help "see" things that would otherwise go missed.
> you get it working without LIDAR and make boatloads of money
And why can't you make "boatloads of money" while using LIDAR? Killing people is bad for business... no?
Before you jump on it, no LIDAR is no silver bullet... but it's available technology that can help quite a bit. Why not use it? Unless you've overpromised what your price point will be for technology that is yet to exist?
[1] https://www.telegraph.co.uk/technology/10567942/Supercompute...
Because Elon wanted to advertise his cars as having all the hardware necessary, because the automotive automation hype is very real. He wanted to sell luxury cars to tech enthused idealists, but he didn't want to cut down on his profit margins by actually delivering what he was selling.
The fact that he always admitted the software was lacking gave him plenty of cover. It allows his customers to continue to believe, even though their car can't actually drive itself. That's the beauty of selling hopes and dreams instead of anything concrete.
I really believe you can make a vision+radar only self driving car, but not without a lot of compute power. LIDAR just makes it easier. Will HW3 be enough? Probably for a lot of scenarios, but it's definitely not going to be level 4/5 anytime soon.
1. using that logic we would still be trying to build airplanes with flapping wings like birds.
2. humans aren't that good of a drivers.
[1]: a la https://www.youtube.com/watch?v=6QCF8tVqM3I or https://www.youtube.com/watch?v=o02H2xGIecc
No they won't. They have all the issues that pure visible light sensors fundamentally have in Earth weather and lighting conditions. There hasn't been much choice for human drivers obviously (though in principle there we could have pushed for better HUD ages ago, and could still make use of AR), but with the switch to computers we can and should do better. It is foolish not to make use of better data input.
>We are an existence proof of that.
Proof of what? Millions of deaths and casualties every single year due to driving? We tolerate it because the benefits are even bigger, and there aren't any other options. It is literally worth mass suffering and death, not to mention economic expense, to gain arbitrary point-to-point mechanized transportation. But that acceptance is purely a relative matter because there is nothing better, as with medical technology standards can and will change once we can improve. There is no reason that self driving cars shouldn't generally be able to see animals (deer, moose etc) in a pitch black night in the fog for example and keep from running into them. Yeah that'd be impossible for us, but that's not a law of physics just a limitation of information input limited purely to the visual spectrum. Using humans as a standard is really foolish given how objectively terrible we are at this.
Yeah, it would be foolish to use my grandmother as a benchmark for driverless cars. Humanity would experience a net increase in safety if she were to take a driverless car to church.
But even the best driverless car today with all their LIDARs can't outperform my father, a 40 year veteran of UPS with literally millions of accident free miles under his belt in every driving condition imaginable - rain, sleet, snow, fog, bright sun. All with binocular cameras, and a finely tuned driving ethic. If I presented UPS with a fleet of similarly capable driverless trucks, I would be a billionaire, and the world would be net-safer.
[0]: https://en.wikipedia.org/wiki/Motor_vehicle_fatality_rate_in...
[1]: 20 mph average driving rate (guess on my part but they stop a lot) * 8 hours a day * 5 days week * 50 weeks a year * 40 years = 1.6 million miles
Over all human drivers we don't get that result, of course. But how many were following too closely? How many were distracted? How many were sleepy? How many were intoxicated? How many were going too fast? How many were old? How many have a slow reaction time? How many weren't wearing their glasses? All of the above factors have nothing to do with the sensory input to the controller, but are indictments of the individual controllers themselves.
Some human controllers are better than others. Some people hit parked cars. Some people go millions of miles without hitting anything. They both have binocular input sensors. If we equip a car with binocular cameras, that doesn't mean we will have the same result as humans in aggregate. Computers will never get drunk. Computers will never be distracted. In the limit, where we have the right cognition, I believe we won't need LIDAR to achieve better aggregate results than the ones you quoted. It would be as if everyone drove like my father, rather than grandmother.
What do you think the fatality rate over all of UPS, Fedex, and USPS are? Some quick Googling shows that UPS alone logs over 3 billion miles per year for their fleet, with only 25 deaths per year. That's 0.8 deaths per 100 million miles. So we can see with proper training, a human controller can be 100% more effective at preventing death than the general population. And the question still remains -- why were those deaths caused? Were they due to being limited to binocular cameras? Or were they something we could program away like being distracted or driving drunk? Or maybe some were unavoidable under any circumstance? I don't know, but it definitely shows me that this idea that humans are de facto terrible at driving and therefore binocular sensors are not enough for driverless cars (again in the long run) is questionable.
- Passive sensors have huge range, power and deconfliction advantages
- The better system is passive sensors in superhuman wavelengths
- Even Levandowski's come around to Elon's POV on unnecessary lidar
LIDAR has a lot more "false positives" if I remember correctly, and a lot of the work is going toward filtering LIDAR data and cleaning it up in cars that use it. I could see not wanting to spend the time and effort in trying to tame something like that when they have something working pretty well right now with different kinds of sensors.
Americans kill 40,000 people a year driving with their insufficient binocular vision. It's obviously not enough.