New facts and questions on Uber robocar fatality


Update: More information in following posts, particularly impressions of serious possible errors by Uber.

As expected, yesterday's fatal accident with an Uber robocar has created a great deal of buzz and controversy. There have been many updates since I wrote yesterday's post, and I have updated the article with most of them. My biggest question, however, revolves around the police statement that the victim was crossing from the west (left side) but the debris is in the right lane, at about the place where the right turn lane is expanding away from it. I asked the police spokesman to confirm that she was going west to east and it was confirmed.

As we also saw, the right grille of the Uber vehicle is dented. The Uber was going 40mph on a 45mph road (original reports said a 35mph road.)

The big question is, it seems that the victim had to cross three and a half lanes to the point where she was hit. Two left turn lanes, the left lane of Mill St. and finally half of the right lane to where she was hit by the right side of the Volvo. As a reminder, here is the location on StreetView.

This is quite puzzling. Uber's Velodyne LIDAR should have seen her very clearly, and observed her for almost 5 seconds if she was walking, about 2.5 seconds if she was running. A bit less if she was riding the bike but the police have seen the forward cameras of the Uber and indicate she was walking. In addition, the safety driver should also have noticed her, but that's less interesting, being just an ordinary human mistake.

(I say 5 seconds because the range of the LIDAR is about 100m, and the car was going 20 metres/second. So it would not have seen her until she was on the road.)

While there are trees in the median which the victim may have emerged from, once on the road she should have been very obvious to the LIDAR, which sees slightly better in the dark than it does in the day. Cameras would also have seen her in the bright streetlights but it's easier to understand issues there.

It is pointed out below that it's less than 80 feet (the stopping distance at 40mph) from the debris to where the Uber parked. That means the vehicle either did slow, or the Uber moved, or the debris was released from the victim's bags halfway through deceleration.

I hope Uber can shed more light on just why, according to police, the vehicle did not even attempt to slow down.

Other news revealed yesterday:

  • The road median has paved walkways but also has signs telling pedestrians not to cross and to go to the main crosswalk. As such, under normal Arizona law, the car had the right of way.
  • The woman is indeed homeless, and was recently incarcerated. However, she has family and friends who are aware events. There is no sign of impairment of either the pedestrian or safety driver.
  • The new Arizona self-driving car law brought in just last month may allow Uber to receive criminal liability -- if there is such liability, of course.
  • Police are not saying a lot, though they have made some statements to confirm that pedestrian crossing at that location is illegal and the car had right of way.
  • At the same time, what a terrible street design -- big paved pedestrian paths that go to the beach in the middle of the median, leading to road signs telling pedestrians not to cross the road here! Who would not cross, in spite of such signs?
  • The Uber safety driver reportedly served 4 years for attempted armed robbery, but was hired under an Uber program to give ex-cons a new start. The criminal record need not have bearing on her driving ability of course, however I do believe that Uber's methods for selecting, training and testing its safety drivers will see some scrutiny if there was fault for the safety driver.

On speeding

The original police statement said the car was going 40mph in a 35mph zone. It turns out to be a 45 mph zone, so I have removed notes on whether robocars should speed. The argument for that can be found on my web site.

On safety records

Uber now has a tall order to claim they have a properly safe robocar. Human drivers have a fatality about every 80 million miles, though a bit more frequently on city streets. Killing a pedestrian is much less frequent. This is regardless of blame. Uber now will have an almost impossible job of stating their vehicle is better than human drivers, even if they drive their next 100 million miles without another fatality. Of course, every week there is a new revision on the road, and the software that was in this car will probably never run again.

Uber's strategy

I outlined likely strategy in my article on Robocar accidents. First, they have probably already fully examined the accident in their sensor logs, and know if they made any mistakes or were at fault. They may even know what caused any fault, or will soon.

If they decide they are at fault, if they are smart, they will immediately admit it and make restitution, unless police attempt some sort of criminal liability. If they don't think they have that, they might decide to fight it.

They will also consider the potential of a wrongful death lawsuit from the family of the victim. As a homeless person, actual damages in such a suit would probably be much lower than in a suit for somebody who was actively engaged with their family, but that will be for those looking closely at the situation to decide. Certainly there should be no shortage of counsel willing to take such a case, because it will be precedent setting.

Unless Uber feels it will certainly win and come out looking good, they will probably make an immediate and generous settlement offer. While the counsel may want the famous case, their duty is to recommend acceptance of a settlement that is truly generous. Uber would likely only let this go to trial if they thought they had a very solid chance of winning and being clearly in the right. From a public perception, fighting the family of a woman they killed is never going to look good for them.

If Uber's examination finds the system failed in a major way, that's even more reason to accept judgment and offer a generous settlement. While it's unlikely, if the cause of the failure can be shown to be negligence, the cost can be very high. While punitive damages are actually fairly rare, and require a pattern of deliberate disregard, they are not off the table in a case like this. While Uber has been trying to turn into a new company, their reputation is far from stellar on these points.


You can combine the first and last bullet point of section one. The sign says "Don't cross here" (or whatever) but what it _really_ says is that there are extremely compelling reasons to cross here. It reminds me of the fact that "Share the road" means "The road is not shared well here".

And so did the police! I will correct it.

If the right fender is damaged, unless the cameras show another thing, the woman came from the right.
How the bike is in that place? somebody took it to that place?.
If the car and bike were not moved AFTER the accident, so the SUV braked before the kick. It takes around 20 meters to totally stop after braking. The SUV did not kick the bike with his front part. Is not damaging in that way.
What may be Happened.
The woman was walking or driving on her bike in the sidewalk. Any driver or robocar, will do nothing in that case. Nobody will stop his car because somebody is walking in the sidewalk. Suddenly the woman turned the bike or was going to fell down and turned entering on the street, just in the front/right side of the SUV. The SUV stopped, but needs 20 meters at that speed . The SUV and bike kicked each other.
Better turn left instead or besides brake? may be, but the bike would kick the car anyway, may not in the fender but in the door, but stopped was mandatory in any case.

The physical evidence we see does strongly suggest coming from the right. It makes more sense the car could not see her if she leaped out from the right. But it is possible to hit her on the right side of the car even if she's coming from the left -- but even harder to not see her. Yet police have watched the video from the cameras in the car, and keep saying she came from the west. Maybe they confused west and east?

And yes, the SUV should have needed 80 feet to stop with full braking, but it does appear to be less than 80 feet from the debris to where the Uber stopped. It is possible (the video would show) that the bike was being pushed by the SUV for some distance and then something happened in the middle of breaking, popping some of her bags, to disperse the debris.

And they confirm, she was going from west to east. So I don't see how she was out of the blue with 3.5 lanes to cross before impact in the right lane.

I believe you, Brad, but how can know the police from where she was coming?. Nobody was there at the moment of the accident and for sure was not the driver who saw the woman. The police even did not know the max speed of that place. Do not believe until I see the video from the car camera.

The police have seen the forward video from the camera in the Volvo. They seem quite confident of this.

"...was going 40mph on a 35mph road..."
Using the handy map link you posted, I just went down the street a bit to the bridge and just before entering the bridge is a 45mph sign. What's really crazy is check out the sign immediately before the accident; it says "Begin right turn lane - Yield to bikes".

Even the police statement said it was a 35mph zone. Uber would have built a map of the speeds and probably started accelerating at that sign.

The new claim is that the woman stepped right in front of the vehicle.

Actually, that was an older article, released before the press conference. But it's the police chief, rather than the police media relations officer. But the official police statement is: "The vehicle was traveling northbound just south of Curry Road when a female walking outside of the crosswalk crossed the road from west to east when she was struck by the Uber vehicle."

But as my article states, it is baffling that the Uber could not detect her if she had to cross 3 lanes.

> But as my article states, it is baffling that the Uber could not detecte her if she had to cross 3 lanes.

That's what I have been wondering as well: I she had to cross three lanes, there was enough time for the Velodyne lidar to detect her and the internal tracker to set up a track of her movement. Based on this data, the scene prediction should come up with two trajectories (the uber car's trajectory and the pedestrian trajectory) which eventually collide. Such a scenaro, which leads to a collision in the near future, must have severe impact on the car's path planning and lead to either braking, lane change or another reaction to avoid the collision. I really wonder about the software's capabilities in this field. Based on this incient, it doesn't look very sophisticated.

This is the most basic level of sophistication for an urban street car. A highway car, not expecting pedestrians on the highway, might not be very advanced about this, but no way a city car does not have this basic level of ability. Something is not right. Possibly just some fact we don't know or the police getting the direction of the victim backwards. But if not, possibly some flaw in Uber's system, and in the attention of the safety driver.

The police have said in reviewing the video that there was no time to react. That makes me wonder if they have confused west and east. A video of a pedestrian crossing 3 lanes before you hit her would not be described in that way.

I read again one of yours posts. The police said moving from west to east but not that she crossed the whole road from W to E.
She was walking in the right pathwalk, may be she just put half of the bike inside the road, from the right side of the road, and then get bacwards, moving, yes, now from west to east. In that case not time to stop, because the car saw the woman just in the path walk and no reason to take any action at that time.

What do we know about traffic at that time of day?

If she was crossing from left to right in front of the car, and first crossed three lanes of traffic, I have to conclude there were no other cars there.

Do self driving cars not know how to swerve?

It was 10pm. Nothing has been said about traffic, so I presume there was none. Also, anybody there would have been a witness and would come forward or talk to the press, probably. Their licence plate would be on the Uber's sensor logs.

Actually, no, many self driving cars do not know how to swerve. Unplanned swerving can be dangerous. Perversely, if you don't swerve and hit a jaywalker, you are not liable. If you do swerve and cause another accident you could be. The advice given to drivers who encounter a deer on the road is not to swerve, because almost all human injuries in that situation come from swerving. Of course, that is not the advice if you encounter a human on the road but the same logic applies. The car does not want to go into "uncharted" territory.

However we're pretty sure that wild swerving is not inconcevable with Arizona Uber Volvos. I'm thinking of last year's crash.
Did they overcompensate for that? Now the car doesn't swerve aggressively? I get the feeling that the car this time was totally oblivious to the entire problem.

Tempe Police have released a 22 second-long video clip from two of the Uber vehicle's cameras: the forward-facing camera and a camera showing the safety driver.

The forward-facing camera clearly shows the pedestrian pushing her bike, crossing from left to right.

Clearly Uber's software failed to react correctly given the situation. As Brad has explained, the LIDAR data should have seen the pedestrian several seconds before the impact.

Add new comment