Is there really a race to build self-driving cars? It seems that Waymo is eons ahead of everyone else.
While Tesla and Uber have both just recklessly (imo) jumped in and started setting loose self driving cars and making bold claims, Google/Waymo has really taken a slow and measured approach and given great care to making sure their cars are actually safe.
The problem comes when the Large Print promises things that the system does not do.
Another problem came from the implementation of the system where Tesla did not verify that the driver maintains hands on the wheel.
people have died using teslas autopilot
Either way there's not nearly enough data to draw conclusions from.
Just to clarify the downvotes.
Highway driving is safer.
(Disclaimer: I work for an Alphabet company.)
There may be ways to avoid some of the other types of accidents, but those ones (which have happened several times) are difficult to escape.
The real question is what's the safety threshold - safer than median human? 90th percentile? 99th?
The same way defensive (human) drivers have been doing for a long time: leave room between you and the car in front of you. Gauge cars coming up behind you and roll forward if they need extra room to stop.
Does not prevent 100% of rear-endings, but a large portion of them. Most rear-enders are not cars plowing full-speed into you, it's someone misjudging their braking and not being fully at-rest when they needed to be. Leaving some margin for this kind of error helps avoid the whole incident.
I'd be curious to know if Waymo vehicles experience more, less, or the same number of not-at-fault accidents per mile driven as human-driven ones.
Also, there are some network effects at play here, the higher the percentage of self driving cars the more rare that situation should become.
Autonomous cars really have to be able to handle these scenarios at least as good as, and ideally better than, humans: there's not a complicated split-second balancing act to worry about.
> I don't think there was a single death directly because of autopilot (not 100% sure though).
There was a death, and it was very prominent.
That is more cautious, but that doesn't, on its own, suggest better tech. That is that the Google cars are actually safer, more advanced, etc.
Those are the Koala cars. Before those, there were the Lexus which went in highways too.
https://www.theregister.co.uk/2016/07/28/tesla_autopilot_dea...
Even though what they released isn't a self-driving car but a glorified cruise control. I personally don't really see how that is reckless at all.
Given that the technology is useless if you use it as 'intended' (Be aware, and in control of your vehicle 100% of the time[1]) I don't think this debate will be settled anytime soon.
[1] My 97 Avalon drives just fine when operated under those conditions.
Here's an article which describes a 40% drop in the accident rate:
https://electrek.co/2017/01/19/tesla-crash-rate-autopilot-nh...
The NHTSA's figures show a drop from 1.3 crashes per million miles before Autosteer to 0.8 crashes after Autosteer.
Would any of those crashes have been fatal? I don't have the numbers to answer that but I posit that introducing Autopilot has prevented more deaths than it has caused.
However, in a society where people can observe the actions of others and form motivations in response to policies, etc, you'll find that because society reacts fairly poorly to organ harvesting, because organ harvesting is implausible to do at scale without extra bad things happenning, etc, the utilitarian solution is actually not to go about doing it.
Only a naive utilitarian wouldn't try to also remain consistent with something like a Kantian imperative of global self-coherence.
Now, as for cars and testing self-driving on real folks, well, this may be something where the water is pretty murky. I think that society will react poorly enough to early bad events in self-driving that a measured approach is actually the best for saving lives in the long term.
The critique here shouldn't be that "well, utilitarianism sure looks good from afar, but would you murder your neighbor?" It should be "The problem is too difficult to address with utilitarianism because it involves complex societal factors and responses."
> Tesla has not done a good job educating their customers on the limits of their "Autopilot" system
The death was preventable, even with the technology deployed.
Their current implementation doesn't seem to be learning any higher level behaviors (there's a Youtube video of guy using it in a park where it repeatedly accelerates towards islands in the road and then flails off the road as it goes too fast around the island).
Imagine if the video didn't have the friendly European accent and patches of green. Imagine if that was a real life scenario and the greens were a hazard.
I wonder if there are more such videos that show Tesla Autopilot freaking out. It would be an eye opener, Tesla's stock price is dependent on Autopilot among other innovations.
The problem seems to be that Tesla's system is extremely dependent on a nice clear white line at the outer edge of the road. On this road, near the traffic islands, the roadside grass sometimes overgrows the road edge and obscures the white line. The white line has also been scuffed by cars near the edge of the road. [1] The Tesla runs off the road in that situation.
There are two kinds of self-driving. One came up from the DARPA Grand Challenge, which was off-road. That kind first looks at the terrain and obstacles, and figures out where it can physically drive safely. There's LIDAR profiling of terrain. Then it looks at road markings and figures out where it's supposed to go. If the road markings lead into an obstacle or drop-off, it will stop or go around the obstacle. That's Waymo.
The second kind came up from lane-keeping and auto-brake systems. Those are very dependent on highway markings, and only work right on freeways. That's Tesla's original system.
It's hard to tell about the others. Volvo has an extensive sensor suite. Otto seems to be mostly a lane-keeping system for freeways. Uber hasn't released much detail.
It may have gone the other way. Remember, the primary sensor technology in 99.999% of cars out there is two front facing visual-light cameras pointing the same way on a swivel and couple of mirrors.
There's little evidence that anything Tesla has deployed is an important step towards viable self driving vehicles.
They are apparently mapping the locations of odd situations, but that is pretty meh (and doesn't require the mapping system to be in control of the vehicle).
Ethic issues arise when the expectation is that the occupants will live and then your product falling short on those critical constraint.
Because when you design something for the masses, you are directly responsible for the deaths caused by your product.
This is not a matter of philosophy, from one engineer to another, it's our duty.
There's no evidence to suggest that is the case. If anything, Tesla has made it more dangerous to assume that it's safe and falsely autonomous.
When you create new drugs, the FDA definitely doesn't look to kindly on people who think like Silicon Valley. "Killing people is product testing" won't fly in pharmaceuticals.
I'm glad that we don't have people like you designing drugs for the masses!
Not sure you can attribute that thinking to any of the actual engineers working on this.
Is there any evidence that these companies haven't gone through closed circuit tests and passed before testing in public roads?
The next step in drug trials is, in fact, testing on humans and some do suffer serious injury or death. I don't think there's any way around it and the greater good of having a pharmaceutical industry at all outweighs those unfortunate incidents.
The outfall is that it led to bad PR for self-driving cars.
Bad PR? Example? I haven't seen much negative spin on self-driving cars in mainstream coverage recently. If anything it's been highly optimistic while still being honest by mentioning the difficulties the companies are trying to overcome.
I don't think people are under-estimating the challenge here, even the layman. But the potential ROI when it does work would dramatically be safer, environmentally friendly, and efficient with human time.
Given the current state of driving is highly dangerous, we'd benefit from iterative progress towards self-driving cars, ala what Tesla is doing. Controlling risks doesn't have to mean holding tech back until it's perfected.
33,000 people die each year on US roads and self driving cars offer the chance to dramatically reduce that figure over time. The more aggressively we can test self-driving software now, the faster the software can be improved.
So as long as the accident rate for autonomous Teslas is initially no more than for human drivers, the Tesla approach will lead to fewer deaths in the long run.
It's unlikely that Tesla gets much useful data from vehicles. To debug a vision system, they'd need the camera data from all the cameras, and that's too big to upload over the cell phone data link.
[1] https://www.forbes.com/sites/brookecrothers/2017/04/23/tesla...
They just need to send back data from a few seconds before either of those scenarios to quickly accumulate a giant library of one-in-a-million edge cases they can test future algorithm tweaks against. I think it's a reasonable strategy.
This article summarizes the latest AP2 software release: https://electrek.co/2017/03/30/tesla-autopilot-2-0-camera-8-...
Do we know how many lives have been saved from Tesla's autopilot? I've only seen anecdotes but I get the impression that it's already a big net win for safety.
For example, nuTonomy, the start-up at which I work, has deployed a similar trial in Singapore in August 2016. (btw, we are hiring in everything!).
It's not a "race", it's more like a "marathon": there is a big difference in making something that works 99% of the times (sufficient for a trial like the above), and something that works 99.9999...% of the times (a product that can be actually deployed).
Namely, that Google relies heavily on their mapping services to make their cars work. This makes a lot of sense for Google, because of self-driving cars require their mapping data, there's a big new market for them. The maps they run on are significantly more detailed than the public-facing Google Maps/Street View. They work in Phoenix, AZ, Mountain View, CA and Austin, TX and like... nowhere else. (Was there one other city in Oregon maybe?)
As an aside, note the "As an early rider, you’ll be able to use our self-driving cars to go places you frequent every day, from work, to school, to the movies and more." I am curious if this suggests you have to tell Waymo in advance a set list of destinations they can ensure work correctly or something similar, or if I'm reading too much into it.
Tesla, Comma.ai, etc. are not relying on special map data as much as they are relying more heavily on road signals and lane lines and such, and then having machine learning decide how to navigate them locally.
While they may not have the same driving record, everyone else's approach works nationwide (and wouldn't be exceptionally hard to extend to a global scope, presuming you taught them different countries' markers and signage), whereas Google's approach currently does not scale.
There is also the possibility that having a detailed map accelerates progress (better automatic scoring of sensor based modeling systems).
Do you want a system that only works on forward facing vision? Or only lidar? Or only gps and road databases? Or only machine learning? We know all of these have holes and blind spots, and a safe system wood have redundancy.
I think the actual distance between Waymo and those that evolved from lane keeping is far bigger than people realize and the biggest danger is people shipping MVPs that kill someone or run into a school playground, if that happens, congress will ban these cars or regulate so heavily that progress will be slowed.
Bear in mind, with this announcement, they aren't even committing to "it works in the Phoenix metro area". There's that "parts of" statement, that indicates that only parts of the area is mapped and hence the cars are only capable of operating in parts of the Phoenix metro area.
If you have evidence to the contrary, please, by all means, feel free to share.
How would Waymo possibly build a car with Level 3 or 4 capability if they weren't doing a hellava lot of onboard visual field processing? Accurate maps data won't let you deal with bicyclists or pedestrians properly, it won't deal with all kinds of hazards.
There's a reason Waymo cars are festooned with LIDAR and cameras and other HW in a giant boxy minivan, and it isn't purely for show to make them ugly on purpose.
I am saying there is no evidence, or even a claim by Google, that their cars are capable of functioning outside a pre-mapped area. You have presented zero evidence of it whatsoever, and are trying to use Google's marketingspeak about how fancy their machine learning things are to insinuate I should assume Google's technology is more advanced than they can demonstrate.
You are highlighting particular capabilities of their collision avoidance on that mapped area to suggest they don't need a mapped area, which is also a non-supported claim. It is entirely plausible, and in fact, likely, that Google's visual mapping for collision avoidance is sophisticated, and yet still fully dependent on a map as a baseline.
Google's marketing claims continue to be extraordinary, and often untrustworthy, and you're going to have to do better than that.
Find me any, literally any, evidence that a Google self-driving car can operate outside the very tiny fenced in areas Google says they can operate. As far as I can tell, there is none, because you are making a claim that even Google itself is not making about it's cars.
What does it mean "cars are capable of functioning <in an area>"? What does "Functioning" mean? Tesla Autopilot don't function off highways, and barely function on highways. So you think someone should get credit for functioning at scale everywhere, if they ship a consumer product with no restrictions, but it fails badly when people actually try it? You're comparing a broken product to one that purportedly works, but you think it's smoke and mirrors?
I love Tesla cars and I plan to buy a P90D, but in my opinion, their whole self driving program is incredibly reckless. (https://www.youtube.com/watch?v=fQxIhMBKblY) Waymo's approach is careful, over engineered, defense-in-depth. Slow by Silicon Valley standards, but you're dealing with public safety. Yes, they use maps as one of their sources of truth, they'd be reckless not to, but they also use LIDAR and vision systems with machine learning, because of course, no map can be real time.
Tesla is trying to sell an upgraded lane-keeping system as a self driving system. Maybe you should be more concerned about that.
I can only say what is public already, but you can look at the disengagements data to see Waymo cars are three orders of magnitude better than their competitors.
If the lane lines are hard to read in a part of Phoenix, Google can ask for them to be repainted before approving the cars for that area but everyone else just has to assume bad lane lines are something they'll contend with.
Is there an NDA these early riders will have to sign?
Do they have to provide a list of destinations ahead of time?
I feel this announcement makes it seem like these cars are ready for public use. But I don't see the evidence to suggest they really are. And as you indicate, Googlers aren't talking.