Tesla Autopilot HW3 details
reddit.com
reddit.com
There was a hackers forum where daily discoveries were discussed.
It was indeed exhilarating, and exactly that way you describe it: on one hand, we didn't really want the thing to be cracked, but on the other it was impossible not to be rooting for those scrappy hackers going down the wrong path at first before figuring out the right one, one step at a time. Every morning, we'd log in to that forum to check their overnight progress.
Still, we were confident that our bank-strength crypto algorithm would prevail.
It did not. :-)
While we had done our due diligence, an external implementation partner had decided to change the audited code later in the process, which broke things completely, in the most embarrassing way.
In the end, it didn't matter, and we probably sold a few more units that what we would have sold without the broken crypto.
Fun times.
I wonder what the "most embarrassing way" would mean in this context -- I'm thinking timing attack or padding oracles, but it sounds like it might have been even more trivial.
There really is an XKCD for everything...
There was a CAPTCHA used to prevent bot spam on a contest entry portal. The code that randomized the image displayed was modified to be stored in an application cache that persisted across sessions (meaning every "user" saw the same image and could use the same answer). Guess how useful that was in preventing bot spam?
(The fix was to delete one or two lines of code that were not only not helpful, but obviously harmful!)
It was summer 2001. The 3Com Audrey internet appliance wasn't quite canceled yet, but it wasn't thriving in the marketplace, either. I picked one up for $50 from uBid and started poking at it, blogging about my discoveries along the way. I figured out that the OTA image checksum was relatively simple to generate (something like two's-complement addition of running 32-bit integers that needed to sum to a specific constant). This enabled me to alter the next OTA image that came off the wire, which allowed me to begin replacing resources and binary components in the system.
By this point people were following the blog, occasionally linking to it on Slashdot. But then something strange happened: I started receiving anonymous emails from individuals who knew a lot about the Audrey OS's internals. They gave me hints about where the more vulnerably coded parts of the system were; they suggested hard-to-stumble-upon but valuable URLs in the Marimba OTA system; and one day when I foolishly overwrote the bootloader with a malformed image, just a couple days later a brand-new Audrey motherboard mysteriously appeared on my doorstep (at the time I didn't live very far from 3Com, and this was an age of the internet when nobody had much reason to hide their identities).
Thus, with the help of numerous knowledgeable individuals, I was able to bootstrap the Audrey modding community. The focus of the world changed soon thereafter, one Tuesday in September, but it was otherwise a similarly exhilarating experience to the one TomVDB describes. The difference was that I was the hacker, and the opposing team -- the one that had built this amazing machine (which I still have two of, by the way, both unopened) -- actually wanted me to succeed, hopefully giving the Audrey a better chance than it got during its initial, ill-fated, mid-dot-com-bust launch.
edit: sounds like efforts shifted to pirating internet streams. Found this neat story: https://www.fiercevideo.com/cable/nagravision-wins-101m-pira...
I remember showing a couple friends of mine how to flash cards with a DB9 serial port card programmer. You could get every channel, including PPV and porn. I never wanted to do anything illegal except watch TV and movies, but a couple friends of mine had highly successful "satellite TV installation businesses" where they would install for legitimate DirecTV customers, then say "you know, for a couple hundred bucks I can unlock all the channels..."
It was a fun time, and what made it interesting was the cat and mouse game that DirecTV played with the hackers. We had cards working for months, but at some points in time you had to get new updates every few days to keep it working. Eventually it became impossible, but it's been over a decade so I can't remember what the final event was.
One of my favorites was when a reaction wheel failed on the Kepler spacecraft in 2013. Someone on Reddit declared that the mission was over with no chance of recovery. I kept my mouth shut. But I knew people down the hall from me were working on a solution. Kepler ended up observing for 6 more years.
Which forums would you say are worth lurking in? This one? The Stack Overflow family? Are any specific subreddits not-terrible?
I've seen some good discussion on Stack Overflow, but I don't regularly lurk. I do lurk on satobs.org - occasionally there is very interesting discussion there.
The most click-baity "ten reasons why you should use my favorite JS framework" posts seem to win, and it's troubling. We will soon be a world full of dumbed down workers, who can't experiment and think for ourselves, but instead get advice from whatever click-bait article included enough buzzwords to get recommended by an algorithm. What ever happened to tinkering and figuring it out on your own?
I work in automotive software development and the same holds true there. ;)
The TPU V1 is a 2015 chip (publicly announced in 2016, and that Google had been using internally for over a year). 4 years is a big lag in terms of technology.
Additionally, the comment about "actually run neural networks " seems to be, AFAIK, plainly wrong. The first version was limited to inference because it could only perform integer operations, while V2 and onward support both training and inference, as they can perform float operations.
TPU V1 was never available to the public, but both V2 and V3 are available on Google Cloud. I don't have any info on it, but at this point, I'd expect all V1 chips to have been deprecated and likely removed from production, given the savings in space and power from V2 and V3.
Well, Tesla clearly isn't doing safety critical work on autonomous systems (it wouldn't have even launched "autopilot" if it was), so 4 years is indeed a big lag.
Isn't then the safer thing to release and encourage autopilot to the largest extent possible?
[1] https://www.tesla.com/blog/q3-2018-vehicle-safety-report?red...
Plus, the statistic ignores the fact that responsibility is actually shared between humans and the car. If autopilot gets into an unsafe situation and gives control to the human at the last moment, the car company might claim that the accident happened under human control. We have a long way to go before we can truly understand the safety of semi-autonomous vehicles.
We have lots of experience with the reverse, where the human drives but the autopilot rips control away when things go wrong. Traction control, radar-triggered braking, even basic ABS is a machine taking over from the human once things get hairy. Several AirForce fighters have systems that will sometimes ignore the pilot's inputs and save the aircraft from a crash. That's the real path for semi-autonomous cars imho: you do all the driving until the robot steps in to save your ass.
(I cannot wait for the day someone tries to sell a car that will automatically obey speed limits regardless of driver input. That's a stock to short.)
Controlling for all those variables is hard and not to their benefit, thus I'd never expect their pr team to do so. But I wonder if that might even creep into their automated driving stats... e.g. by drivers taking control in dangerous situations.
Why would you do so?
This is like saying "Let's assume the sky is purple. Then, we can conclude that the sky is purple."
For both software and hardware there will be requirements for documentation, design, verification, and testing. Some even go as far as to requir the implementation of certain functional logic in multiple ways and having concensus logic (eg hardware logic to interpret GPS accuracy/confidence That ends up being broadcast externally).
There's also DO-160 that's concerned with environmental requirements (eg temperature, humidity, lighting).
This is what makes a radio that should cost $300, cost $5000, and why FAA certified aviation components are expensive compared to uncertified components (eg hobby airplane builds).
One slight nitpick is that DO-178B or DO-178C are revisions of the standard and the associated criticality is associated with the Design Assurance Level (DAL) from A to E. One would say software is developed to DO-178C DAL B for example.
The DAL is determined before the software is started at the Systems level according to a process described in ARP4754A (Guidelines For Development Of Civil Aircraft and Systems) looking at the system architecture, hazards, and potential mitigations of those hazards, one of which is developing software to a certain DAL.
As software people I think many will laugh at the low TRL of their own work. It isn't too bad or anything since other sectors need to move fast (probably security people will disagree). But other sectors need to move slow and ensure that things don't break. Because things breaking means people dying.
[0] https://github.com/kubernetes/kubernetes/blob/ec2e767e593953...
[1] https://steveblank.files.wordpress.com/2013/11/nasa-trl.jpg
I doubt more than 5% of developers have ever worked on anything beyond a TRL4. SREs, maybe,
Once software is operational, whether you are following DO-178x/33x, FDA's General Principles of Software Validation and other guidelines, ISO-whatever, or NASA/military TRL, there's a whole ton of stuff you have to do to make any changes once you reach a certain point in development. Since most businesses have the benefit of captive employees that can be woken up in the middle of hte night and terra-locked systems, they never even come close to reaching that point.
Same reason why New Horizon used MIPS R3000 CPU in 2006, where as it was also used in Playstation 1 in 1994.
It would be very interesting if Google was selling Tesla EOL TPU v1 units. The specs are very good for what Tesla is trying to do.
4 years is not a long time considering this is a proprietary technology that even GCP customers haven't had access to for very long. It will likely be a few more years before any TPU-like technology is available in the mass market.
... until it mis-view an exit lane and attempts to crash/kill you, that is (see events of 2018).
That is kind of deal breaker for me, when talking about 1+ ton metal box moving me around I will take "slow to market but not blind" instead.
I'm still waiting for Tesla to accept reality and put lidar (make the option more expensive if you have to and you probably do, whatever)
The way Elon Musk tells it, they put it into production once it was safer per mile than human drivers. Which seems like a legitimate point.
You use something when it's better than what you would have to use in the alternative, not when it's 100% perfect with no possibility of ever making a mistake under any circumstances. Which is probably not even actually possible to do.
The problem is if human drivers kill more than 30,000 people in a year, it's not news, because it's been that way forever. But one autonomous car kills one person and it's the top story.
I'd caution you against believing anything that he says. It's easy to extract from the data an unfair comparison that shows the group with Tesla AP being safer, but comparing it against cars in totally different driving context or against the safety of motorcycles is just completely deceptive. The stats indicate that Tesla w/AP enabled on freeways is less safe than Tesla w/o AP enabled on freeways.
But Tesla knew it would certainly be perceived as such.
How? I'm not trying to instigate something here, but a quick search of this isn't showing anything except a bunch of groaning about how unfair it is that Tesla gets all kinds of tax breaks for its cars because it's the only one selling electric at any kind of scale.
https://www.businessinsider.com/tesla-liberty-mutual-create-...
"The firm's InsureMyTesla program is an example of Tesla working directly with an insurance partner to mitigate risks and keep rates low."
I thought subsidize meant that they would be basically paying the difference.
If there's no money changing hands in this partnership, then either the insurance companies are really dumb and are just throwing money away, or there's some truth to the stats that tesla keeps bringing up (40% fewer crashes in cars with autopilot) (or there's something else i'm missing!)
That's all marketing fluff.
The only U.S. insurance company offering such a discount is some random startup insurance company available to only a small percent of the population. The company is about 2 years old, requires significant tracking of your driving behavior, and it's entirely unclear why they are offering the discount.
The only UK insurance company offering such a discount said “it was too early to say whether the use of the autopilot system produced a safety record that justified lower premiums. It said it was charging less to encourage use of the system and aid research.” It made this decision working closely together with Tesla, it wasn't based on any actuarial data.
Also, from that article, it looks like the NHTSA found that crashes were 40% lower in cars with autopilot. [2] I don't know enough about insurance and how it works to know if that counts at all, but it shows that it's not all marketing fluff.
[1] https://www.businessinsider.com/tesla-liberty-mutual-create-...
[2] https://www.businessinsider.com/tesla-autopilot-cuts-crash-r...
"Liberty Mutual Insurance offers discounts for various electronic safety features; such as Adaptive Cruise Control, Adaptive Headlights, Collision Preparation Systems (including automatic braking), Blind Spot Warning, Lane Departure Warning and Rear-View Cameras"
I'd be interested in seeing crash data for Teslas compared with other $50-$130k cars less than 5 years old.
I don't know it well enough to know if there were other factors at play, but it wasn't just that Teslas in general were less likely to crash, only the ones with AP hardware installed.
Insurance companies may do this because it is a strong signal the customer is gullible to questionable upsells. Maybe they've found that statistically they can sell them more things like overpriced umbrella insurance through direct mail marketing pack ins with their car insurance bill.
(edit: opps, I misread AP as FSD)
It's also only used/active when it is reasonably able to work. Therefore the miles used are inherently better/safer as it is (much less whiteout / torrential AP driving).
This is a luxury human drivers do not have, short of "do not drive those miles at all".
I have always hated this metric. It's disingenuous, and selective.
https://www.businessinsider.com/tesla-autopilot-cuts-crash-r...
They may not be taking the hardest miles, but the only way to improve the overall average is to be doing better than the status quo on the miles they are taking.
The thing is that I can take precautions to increase my safety margin - don't drive impaired, drive in good weather conditions, get lots of sleep, don't be aggressive, don't follow closely, etc. etc. etc.
I don't have a Tesla, but I imagine there's nothing I can do to voluntarily make it safer. It functions normally until it doesn't.
It's common knowledge that most people think they're better than average drivers, so we tend to compensate and believe we're no better than the statistical average driver. But yes indeed, most of us in fact are better than average drivers whenever we're doing our best to drive carefully.
In that sense these Level 2 systems are a lot like q-tips. The box says you shouldn't stick them in your ear, but everybody knows that's what you buy them for. If everybody were to follow the directions, the consumer desirability of the product would plummet.
Even though you're not supposed to take your hands off the wheel, level 2 systems are often referred to as "hands off". And during a 60 Minutes interview Elon Musk is seen taking his hands off the wheel. Unofficially the point of Level 2 automation is that you can take your hands off the wheel. Officially, your hands must remain on the wheel. The unofficial usecase is what actually gets consumers enthusiastic, but the official usecase is what companies like Tesla cover their ass with.
They're having their cake and eating it too.
They were comparing crash and fatality statistics against things like motorcycles.
So no "over a billion" isn't plenty.
Are they safer? Possibly. Statistics show they are safer? That's bollocks. There's not enough data to say yet.
There are lots of things wrong with auto-pilot, but it is far from "garden-variety".
is it flawless, no, but it does work in conditions I didn't find fun driving in. An example was two lane highway, at night, in the rain, with enough jokers with badly aimed lights or not dimming their lights. It did very well and let me feel confident looking away from the road if the oncoming car's lights were too bright.
I did not buy the FSD suite, I don't have a desire for that level of autonomy and I am not confident anyone will pull it off soon. I mean I have seen examples where they map it all but that isn't the same.
Here is a video from Cadillac showing your statement is false: https://www.youtube.com/watch?v=_rxW68ADldI&t=0m39s
Tesla will speed up or slow down and change lanes automatically. This is not the same. Also, Supercruise is only enabled in very limited circumstances. You can use autopilot anywhere even if it isn't officially recommended.
You are still required to engage the turn signal; Tesla will speed up / slow down to execute the merge, eventually, and will steer the merge for you but does not make the decision to do so autonomously (though it'll recommend a new lane to you with a blue lane line on your dash).
I would advise you to sample one before drawing conclusions.
Tesla’s do great in traffic.
I swapped in October to a 3PE, now I'm seeing that another upgrade is incoming, and that THIS one is required for FSD.
I've spent plenty of money with this company, but handing over $5,000 more dollars for a feature that is unproven is not something I feel like I should need to do; especially when I'm clearly advertised to at the time that I can buy the feature later for $6,000.
If fully autonomous driving becomes available, and it costs more than what was advertised, they are going to have a big problem.
Can you go into depth about why this comparison is inappropriate, why it unlikely to be solvable in this way?
(I can certainly understand “not now”, but your claim seems rather stronger than that).
We’ve made extraordinary progress, but, ultimately, we’re still training systems in extremely simplistic ways to solve (conceptually) simplistic tasks. Localizing objects (into a set of categories no less) is only a fraction of what your brain does when looking around. Moreover, we still struggle with this. One of the Tesla accidents was due to a segmentation error (two simimilary colored and textured objects were merged).
You'd need a storage rack in your frunk to record any amount of raw sensor data. That's before you consider their production AP hardware doesn't have cycles to spare to do any of that recording.
Source?
> You'd need a storage rack in your frunk to record any amount of raw sensor data. That's before you consider their production AP hardware doesn't have cycles to spare to do any of that recording.
I have no idea on what you base your claims, but without more information it is hard to figure out what you have in mind...
What curious people have found is that the car just sends a basic disengage report with maybe 10s of h264 video or spaced far apart raw images from the cameras.
You don't (always) need full resolution with 30 fps raw frames for training your model though. If you are looking for missing exit, stopping at a traffic light, etc. you need front facing camera (so 3) and a few raw frames (and the sensor is 12 bits not 16). One frame is 1.8MB raw, but if you train on YUV images, this would be even smaller. With a 200MB buffer in memory, you can keep 100 frames at any time (and you have 8GB local storage for storing while waiting to upload)
You said "any meaningful amount of data", I guess this is subjective!
https://electrek.co/2017/06/14/tesla-autopilot-data-floodgat...
There's two tiers-- Enhanced Autopilot (EAP), and FSD. To own FSD, you need to have already purchased EAP. However, you do not need to purchase FSD to have EAP.
Enhanced Autopilot is (in simple terms) lane keeping + adaptive cruise control. It's designed only to work on highways, and does not need the HW3 board nor will owners that only purchased this get the new hardware.
FSD owners will get the HW3 board so that Tesla can fulfill the promise of providing FSD that the owner originally purchased with their car (or after-the-fact). FSD will (in theory) handle self-driving in most cases, such as regular street driving with stop lights, stop signs, roundabouts, etc.
Why should we believe Tesla that HW3 is actually capable of FSD this time?
The things that could happen are that either Tesla ships full self-driving (FSD) and better delivers when it does (i.e. it has to work reliably).
Or it puts off the customers again, promising a new hardware 4, 5, 6, ... capable of FSD - in which case people better read the instructions (and pay attention at all time lest they end up in a divider or under a trailer).
When you select full self-driving capability in the Tesla configurator, it states explicitly and prominently that this isn't enabled (or ready!) yet, and is subject to legal approval.
Anyone that buys a Tesla, thinks it is fully autonomous, and dies as a result is failing natural selection. Third parties (pedestrians and other drivers) on the other hand are victims, but have there been any examples of this with Tesla cars? My not-full-self-driving AP2 Model X can detect pedestrians, and will also emergency stop if it detects things. This is aside from the fact I, as the driver, should be emergency stopping if I see a forthcoming collision with a pedestrian!
Think of Autopilot like a horse. You still need a driver, but at least you've got another set of eyes on the road. Sometimes it will notice things you don't, sometimes it will get spooked by things it shouldn't do, but two sets of eyes/two brains are better than one.
https://www.techrepublic.com/article/tesla-to-release-fully-...
If they don't have FSD by mid-2019, that's two years off that target. Deadlines slip, companies overpromise, this is a hard problem, all that jazz, but by early 2020, people who leased Teslas that they were told would have FSD are going to be deciding what their next move is. If part of the calculation is "FSD still ain't here yet, but Tesla is promising a free upgrade to HW4 which they say will really lay the foundation for FSD this time Elon pinky-swears," I would submit that it's gonna matter.
You'd have to elaborate as of why this would have been the main reason (vs other possible reasons...).
https://www.dmv.ca.gov/portal/dmv/detail/vr/autonomous/testi...
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.
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.
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.
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.
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.
[1]: a la https://www.youtube.com/watch?v=6QCF8tVqM3I or https://www.youtube.com/watch?v=o02H2xGIecc
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.
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
Americans kill 40,000 people a year driving with their insufficient binocular vision. It's obviously not enough.
>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...
Bringing as much of the software and architecture in-house as possible shields them a bit more from that happening again.
I have no idea if that is the ultimate reason, or if it was even a significant reason, but it sure can't hurt!
But it's probably somewhat bad for nVidia given how much effort (and marketing) they've been putting into their autonomous car platform.