Uber settles quickly


Uber has reached an undisclosed settlement in the fatal incident with the victim's husband and daughter. This matches my prediction of Uber's likely best course of action, since it will shut down much of the public discussion and avoid dragging all sorts of details out into the open in a lengthy trial. The settlement comes with an agreement for silence, as you might expect.

Of course, that means the public does not get to see them, at least for now. It will see some of them in the NTSB accident report. If there is a criminal trial over the death, some details will come out, but far fewer than would in a civil trial. If Uber is simply cited for infraction of the vehicle code, it would not want that to go to trial and would just pay the penalties.

Another option lies open. The governments issuing permits for Uber to do its testing may require it to be more forthcoming if they wish new permits. Last week, Uber's permit to test in California expired and it declined to renew. It is unsure who will let them test in the future.

The settlement offer was presumably generous, so that the family would quickly accept it. As noted this is compounded by the fact that society and court cases don't attach large values to the lives of the homeless in contrast with others. While there should have been law firms willing to work pro bono (or certainly on contingency) because of the fame of the case, they would have to be looking at a highly uncertain amount of damages. If the court simply tried to calculate the future income of the deceased and the pain and suffering, it might not have generated a number that would justify the very expensive prosecution of a trial like this. Only damages that are multiplied because of the deep pockets of Uber and far from certain ability to show massive negligence or evil intent would justify that. The family is wise to take any suitably large offer.

Does Uber even know what happened?

There is some speculation, however, that Uber may not have yet been allowed access to the vehicle and its data log hard drives. They might not yet know much more than us. If so, they are probably in a very strained emotional state to be sure, knowing they ran over this woman but not why. The settlement makes this slightly less likely, but the reality is that whether it happened due to a stupid failure or an obscure freak bug, the result to the family is the same, and so the settlement might be the same. The desire of Uber to get the story out of the news is the same.

Uber would certainly have begged the officials to make copies of the disk drives for Uber, and you would think they should, since only Uber can trivially read their own internal log file formats and assist police, but they might have refused to copy them.

Uber will have, like all teams, a checklist of what to do in an accident. Of course, first will be calling 911 (possibly after getting out of danger and if trained, administering first aid.) Second will be calling Uber's own emergency hotline. Uber team members would have headed there quickly, but police and ambulance would be there first. Police might tell Uber staff not to touch the vehicle or any evidence.

Uber would have data if their vehicles are logging over 4G networks. They can't log the full sensor stream, but they could log some distillations and diagnostic messages. But they might not do that at all. If they thought about this, they would also have the car, from the time it has been in an accident until it is shut off, to be sending as much of the full data log it can, back in time, over the data networks. Done properly, the car would do this on its own, knowing full well it had been in an accident. Or the safety driver would invoke a command -- though this safety driver was surely extremely distraught.

A third option, which might be implemented after this, would be for the car to copy as much data as possible over wifi to the safety driver's phone, while also sending it to the cloud.

But perhaps Uber didn't think of such plans. This might explain their silence at a time when they really need to reassure the public. If they don't know, however, then they can't talk, for fear of saying something wrong.

Markets can be cruel

I'm at Nvidia's GPU conference, where a lot of car developers are present. NVIDIA decided, as a precautionary measure, to announce that they would also stop testing for a while. This caused their stock to lose $30/share as investors imagined this meant some connection to the audience. NVIDIA later issued a statement that Uber was not one of their partners. (This does not mean Uber isn't using their GPUs with their own tools, as many teams do.)

Detecting safety driver inattention

While the program was set long ago, some presentations were apropos. Lex Fridman of MIT presented his team's research on using neural networks to classify what drivers are doing. He demonstrated his tools on the released Uber video of the safety driver, and they easily detected she was looking away from the road for extended periods. Their software at his web site has been released open source, making it fairly inexpensive for teams to install tools to track safety driver attentiveness and warn if they are getting drowsy or performing below par.

The use of such tools is actually commonly talked about with "autopilot" offerings that need a constantly attentive driver. There is much less talk about using them to monitor safety drivers. These free tools should make this happen more, and could even open up discussion of laws requiring such monitoring. Possibly it would have prevented this death.


I'm very concerned public safety knowledge and challenging attitudes which may be felt towards homeless people is being covered up by an NDA. This is a huge public loss. Would there be a public interest defence if the family borke the NDA or is it possible a law may be passed to gaurantee that public interest couldn't be gagged?

I also hope the wellbeing of the safety driver isn't forgotten. She may be left with trauma to manage and need rehabilitative support.

One question which just popped into my mind is how would an AI driver car cope with emergencies such as a person having a life threatening incident leaving them needing help. Ordinary people can spot a car being flagged down to help with indcidents like an old person having a fall or a person harmed by an earlier accident, or young women at danger from walking home alone through an unsafe area or even a homeless person stranded by extreme cold weather needing a lift to a hostel or other suitable place of safety. If Uber is to prove itself truly apologetica "sorry" and money is not enough for this to be meangingful. How does this incident now covered by NDA translate into a meaningful change which helps ensure this kind of accident never happens again?

I don't think there's much public interest violated by the NDA, because the family knows only one secret -- the amount of the settlement. If you don't allow families to agree to secrecy on this, it means they have to take smaller settlements, so you are punishing the victims.

The loss of public interest is the fact that there will not be any civil trial discovery. Since you can't really have the law force the family to engage in an expensive and painful lawsuit -- again punishing the victims -- there is no option there. There is still a police investigation, an NTSB investigation, possible NHTSA investigation, and potential action by state DMVs that Uber wants permits from.

However, none of those have the tremendous power of civil discovery, it's true.

These kinds of issues have been raised by UK politicians over NHS healthcare out of court no admission of fault settlements with gagging clauses. This is the kind of thing which informs my view. In the UK there might be a public enquiry or coroners report or police investigation. I don't know what other mechanisms exist.

This article outlines similar concerns by Toyota and Nvidia plus different responses from other names in the autonomous vehicle industry.


Two other AV manufacturers, Toyota and Nvidia, also suspended their AV testing on public roads in the wake of the accident. Neither Toyota nor Nvidia gave any indication they were suspending testing for fear of their own AVs causing similar accidents, though. Toyota said the suspension was out of concern that the accident may have “an emotional effect” on its drivers, while Nvidia’s Chief Executive Jensen Huang said it was out of respect for the victim.

As I said the issue is that the family doesn't know any secrets. It's possible that Uber revealed some details of the cause to them, to show them that they should take the settlement, but they would not do that without NDA anyway. Any time we say, "A settlement should not allow X" what that does is take away one of the levers the victims can offer to increase their payout. And we might decide that's a good thing -- less payment for victims, but more public good served. But I can't see that here because the silence agreement does not seem to bury any major secret the public needs to know. I mean we would like to know the amount of the settlement offers, and that would help other victims perhaps, but also hurt them.

Nvidia paid a price for their shutdown -- the stock dropped $30, but is now recovering. I bought some at the bottom, since I knew nvidia had no role in this incident.

How do you justify heading off potential improvements to regulatory regimes civil discovery and improving social policy? I'm also a bit confused by your suggesting human rights are for sale unless you meant something else? Public interest seems to have a different meaning in the US from the UK and rest of the EU, and perhaps even key markets in Asia. Why should regions with a different staring point be beholden to Wall Street short-termism?

I'm not sure we will see eye to eye on this one so will let this rest and move on.

The public interest is best served by the rapid development of this technology, which is intended to save more lives than almost any other technology in development, with the lowest reasonable risk. That is neither zero risk or extremely low risk, nor is it a completely reckless approach. Most people seem to hope to have a risk within range of the risk of ordinary driving during development, or only modestly worse, and then progress to a lower risk in deployment.

Civil discovery is a powerful thing. It is much more powerful, and gives defendants far fewer rights than they have in other legal proceedings. It is this way because a civil suit seeks only money, not other punishments. But you can't force people to sue, and it's not meant to be an end run around the protections for the accused which are built into the law for a reason.

You can say, "too bad Uber was able to buy them off so they don't sue." But you would be discarding all the thinking about justice which has gone into the system. Of course, one can debate if the goal is justice for the victims at the cost of transparency, more more transparency at the cost of justice for the victims. That's a reasonable thing to debate. We tend to swing towards the justice side in part because it also protects the rights of the accused.

I'm not debating, Brad. I'm sharing a different perspective - an external thesis.

Some issues are not resolvable via adversarial courts or debate. With regard to some issues this is why human rights and discrimination legislation was passed because the protections did not exist in the normal course of justice which is a construct based in part in civil law jurisdictions upon conventions and precedent and in some cases basic access to justice. In the UK issues such as slavery (which was not actually a legal thing in the UK which may surprise many) were only partially resolved by 'Somersets case' and the path of women's rights has been equally tortuous right up to the present "Me too" campaign.

I respect your views on technology but believe human rights and social policy are not your strong suit. I don't believe victimising Uber back is helpful but do hope that people with different input take note as they should especially at the legislative level.

But you would be ill-advised to tell somebody who spent 10 years as chairman of one of the world's most respected civil rights organizations that the subject area is not his strong suit. I don't mean to argue from authority or claim perfection, but please stick to arguments, not personal characterizations.

Pardon? Why did you skip past the entire area of the UK being a wholly different market and the given public interest example? If you are mssing this and providing no input on the broader public policy issues how can a different conclusion be reached? I do believe you focus very well on the pure technology but your social policy and international perspective is absent. I'm not fighting you and this is not a "debate" so why be hostile?

I believe it best to end discussion here and let this issue rest. Thanks for your hospitality.

I'm not sure how much, if any, less power they have than civil discovery. At least for air related things "NTSB investigators have the right to interrogate witnesses on demand, inspect files, enter facilities and aircraft, examine the processes and computer data of any party involved in an air crash. Besides these Congressionally-authorized powers, the NTSB can obtain subpoenas and court orders for special searches and seizures of any party who may have relevant evidence useful in determining the cause of the air crash." https://www.avweb.com/news/avlaw/181884-1.html But, I don't know how that plays out here.

A grand jury would also have enormous investigatory powers. Not sure there's any federal crimes that might warrant a grand jury, but I can't think of anything that an Arizona grand jury couldn't get but which civil discovery could get.

I'm not surprised that there was a settlement. Uber likely was willing to pay a *lot* to prevent a public trial, and the plaintiffs would have had a tough time getting a large judgment. In addition to the issue of damages, there was also the issue that the pedestrian would have been found to be partially at fault - Arizona is a pure comparative negligence state.

That said, if the plaintiffs could have found a smoking gun (like an email between Uber employees which showed that Uber knew about the unreliability of the software in this sort of scenario), there might have been significant punitive damages awarded. It's hard to imagine there's been enough time for the plaintiff's lawyers to have done a thorough job searching for some such smoking gun. Maybe there's a provision in the settlement that invalidates the settlement if something like this is discovered later.

But all of this would have factored into Uber's strategy. I don't know the family but presume they are not one of means if their wife/mother was homeless. So they could only proceed with contingent counsel. And counsel would be themselves looking at the odds -- take good settlement now, very quick income, or take the risk of a very expensive case for a lower paying with the unknown chance of a bigger payout if something damning is found. And it has to be pretty damning.

So Uber's counsel would have been doing this math, and offering a number that pushes it so both family and counsel want to settle. Uber may or may not have all the information (they certainly have the emails) but they also know that the case is bad for them. And plaintiff's counsel knows how dangerous it is for Uber and argues the number up, etc. etc. After 10 days they came to a number.

"This might explain their silence at a time when they really need to reassure the public. If they don't know, however, then they can't talk, for fear of saying something wrong."

There is presumably a lot that Uber does know.
1) They presumably know how the obstacle detection was designed to work, and whether it was designed to work under the conditions in this case.
2) They presumably know how the obstacle detection was tested, and whether it was tested under similar conditions. If they have no testing for pedestrians walking bicycles, for example, or no night testing, or no testing at speeds above 35mph, or no multi-lane testing, or no testing when following another car in the same lane, or no testing outside of crosswalks, or no testing with dark clothing, that would be important information to know.
3) They presumably know if any parts of the system were temporarily disabled or degraded.
4) They presumably have access to around 200 similarly configured vehicles to experiment with, to see if the failure can be easily replicated, and what factors matter.
5) They presumably have some access to results of the ongoing NTSB investigation, although presumably with disclosure restrictions.

There is a natural tendency for organizations to want to reassure customers, employees, stockholders, and partners that a given accident is some sort of rare anomaly in an otherwise well-designed, well-tested, and well-operating system. The longer they go without making such a claim, the less likely it seems to be the case.

We know that their system was designed to work in this situation because everybody's has been for years, and their own demos show the passengers a screen with the perception system output on it, where pedestrians are clearly marked.

We know they have driven this particular bit of road very frequently, according to reports, and they claim to have done something like 2 million miles of testing. We don't know the specifics but it's close to impossible that this sort of condition was not tested.

Number 3 only they know, I agree, and I do expect they are doing tests like described in #4 -- except since they shut down all testing on city streets, they can only be doing it on test tracks. They can be testing the sensor suite and perception systems on city streets, including that very street, and probably are, with full time human driving.

As for #5, that's what I don't know. Rationally, the NTSB should have given them copies of the disks and requested their help decoding them. But maybe they have not. It is a supposition to explain Uber's silence at a time when they desperately need to reassure the world. The other explanation, of course, is that the news they have learned will not reassure the world at all.

"they claim to have done something like 2 million miles of testing."

But I don't know what that means. Is this the first pedestrian to have
gone undetected? Or does that happen frequently? Do they even keep
track of undetected pedestrians or pedestrian near misses?

If this pedestrian had taken a few quick steps at the end to make it safely
across, would Uber have voluntarily suspended all its testing because its
pedestrian detection system was proven to have a serious flaw?

"We don't know the specifics but it's close to impossible that this sort of condition was not tested."

What condition, exactly? I'm fairly sure they've never tested a pedestrian wearing black walking a bike at night across a roadway outside of a crosswalk. Maybe adding the "walking a bike" is superfluous, but have they ever tested the system, live and not as part of a simulation, with a pedestrian wearing dark clothing crossing a roadway at night outside of a crosswalk?

How do you even test something like that? In non-autonomous mode, I guess, with a driver and pedestrian with experience working as a stuntman.

Along those lines though, I'm curious how these companies are testing performance in various near-miss scenarios. They really can't, until they get "lucky" enough to have someone do something stupid in front of them.

How are you sure they haven't tested it? I am pretty sure Waymo has, but they are the best in the business.

They certainly have tested their cars on roads with many pedestrians around, in the 4 cities they do testing. Lots and lots of them. Wearing all the sorts of clothing people wear. Day and night. Their software operators are watching the screens to see if the vehicle is correctly picking up the predestrians.

In addition, if they are a well funded team, they have done test track testing with their own people acting out various roles as pedestrians of all sorts. And they also will have made a simulator where they definitely would indeed test this, and everything like it. Waymo has done 5 billion miles in simulator. That's not 5 billion miles of boring highway, that's 5 billion miles where each mile or less is a differenet strange situation. Including anything that you can think of in a few months of thinking about it, and anything they have seen in 10 lifetimes of test driving.

Uber has not done as much as Waymo to be sure. But people report that Uber drives this road a lot, and this road has jaywalkers all the time, people going from popular parties on the weekend at the "beach" to light rail and other locations.

I have certainly seen things like this many times. Because I've been doing it. I cross in that sort of place often, but of course I do it by carefully using both my eyes, ears and brain to make sure that no car is coming.

"In addition, if they are a well funded team, they have done test track testing with their own people acting out various roles as pedestrians of all sorts. And they also will have made a simulator where they definitely would indeed test this, and everything like it."

Are we left to speculate as to whether Uber actually has a test track somewhere with their own people acting out various roles as pedestrians? If they had such a test track, wouldn't they say so? Wouldn't there be videos? If they had a simulator, wouldn't they try to dispel the impression that they are simply relying on random public road testing?

Does the CEO ever take a turn walking on the test track walking out in front of an oncoming car? Do they have some sort of robots simulating people walking out that are replaceable if they get hit?

When aircraft engine makers test their engines for resistance to bird strikes, they don't just say how many miles they have flown without a failure, they actually shoot frozen chickens into the engines on a test stand and video the results.

"I cross in that sort of place often, but of course I do it by carefully using both my eyes, ears and brain to make sure that no car is coming."

Was the Uber actually visible at the time the pedestrian stepped off the median? It looks like the impact point is about 45 feet from the median, which probably took about 15 seconds. At 38mph, the car would have been about 836 feet away, and the road has a curve it it.

At https://www.uber.com/info/atg/ they say they have a test track, though I don't know much about it.

You think Waymo has tested "a pedestrian wearing black walking a bike at night across a roadway outside of a crosswalk" in live conditions not on a test track? Before this Uber incident?

In any case, it's not something I care to argue about. Maybe they have. I'm more interested in learning about *how* they could test that without putting people's lives at risk. Test track testing (let alone simulator testing) isn't the same as the real thing.

Of course is done less than simulator. It's where you want to test the dynamics of real sensor performance. So the test track is just the place you want to test the different clothing, the weather etc.

Some work on test tracks is done with balloon pedestrians and balloon cars. There is a whole industry about this, it's been popular for years with major automakers. However, at low speeds and with alert safety drivers, the test pedestrians are not at significant risk. I have done it.

Simulator is mostly where you test, "OK, assume the car sees -- or assume it doesn't see -- somebody. Now let's see what decisions it will make." In sim you can do things like, "Have a ped cross the street, but have the perception system not see him for a while" and keep increasing the delay until it hits the pedestrian. You then learn your limits of performance. Independently you can go to the test track and find out how well the perception system sees all types of clothing, all heights of pedestrian, all walking angles, all weathers etc.

For simulator, you write down every situation on a road anybody has ever seen, heard of, read about or can imagine. Everything your test cars have seen in their millions of miles on the road, which you recorded in detail. For each of those, you parameterize them. You try the pedestrian crossing 1s later, 1s earlier etc. etc. You try the car going 1mph faster, 2mph faster etc. You try them all, see if you fail anywhere you should not fail.

Ah, I think I see what you're saying now. Logic-wise, the situation was extremely simple. A large object is in front of the car. You should stop. That situation was certainly tested.

Whether or not the sensors can detect a pedestrian wearing black walking a bike at night across a roadway outside of a crosswalk while the car is moving at 40mph (and the lighting conditions are similar to the way they were that night). I don't think that was tested, at least not with the LIDAR turned off if indeed the LIDAR was turned off.

If I'm right about that, I hope that Uber is never again allowed to test its self-driving cars on public roads. I suspect it's never again going to anyway, though.

Absolutely, for their usual sensor suite, detection of a pedestrian, wearing dark clothes, would have been extensively tested. I mean it is tested by Velodyne, the maker of the LIDAR, which provides specifications on the range of the sensor on different reflective surfaces including black cloth. Unless she was wearing Viperblack, but frankly I think they test even that (and indeed have lower range on that) but she also had blue jeans. Trust me, they also test on 3 foot tall pedestrians, who everybody is even more keen to avoid hitting.

Add new comment