Waymo has been involved in a fatal crash involving a pedestrian in Dallas. According to reports from police and Waymo, a pedestrian was crossing Maple Ave. in Dallas not at a cross-walk. The pedestrian was hit by an SUV driver (doing 35mph in the 30mph zone, not uncommon speeding for Texas) and his body thrown over 30 feet into the oncoming lane, where an unoccupied Waymo robotaxi was traveling.
Waymo has indicated that their vehicle was moving at 24mph and detected the pedestrian had entered the roadway on the other side, when it was about 150 feet away, and anticipated he might attempt to cross the street into the Waymo’s lane. It immediately slowed in anticipation of that crossing. According to security camera video, the man and a companion had just left a store, and the victim stepped into the street, and was struck from behind shortly after.
The pedestrian was thrown a distance estimated as 30 feet, hitting the left rear side of the Waymo i-Pace, which had slowed to 5mph. The Waymo then immediately stopped. (The SUV also stopped and remained at the scene.)
The pedestrian then fell onto the roadway where he was treated by passers-by and emergency crews, but died at hospital.
From these descriptions, the Waymo seems without fault under the normal standards of traffic law. In addition, if the pedestrian was thrown 30 feet by being hit by a vehicle at 35mph, a side contact with a vehicle at 5mph is quite unlikely to have been the source of the fatal injuries, certainly not the primary source.
This incident immediately raises parallels to the infamous incident which ended the Cruise robotaxi project. There, a pedestrian crossing against the don’t-walk signal was hit by a human driven vehicle, and flung into the path of the Cruise. After that, things diverge, as the Cruise vehicle settled on top of her and then in a hurried attempt to pull over, it dragged her, but Cruise followed a path of poor communications (to put it mildly) which omitted that part of the story to the press and regulators. This led to their being pulled from the roads by the DMV and their eventual shutdown as a company.
Cruise was at no fault for hitting a pedestrian who was thrown in front of their vehicle, as Waymo is also not at fault, but what they did after was Cruise’s undoing. There’s no indication that Waymo did anything wrong in this situation, nor that they are hiding some mistake, though naturally there is not normally an indication something is hidden until it is revealed. Waymo has been forthcoming to most questions. Nobody wants to repeat Cruise’s mistakes, both on the road and in communications.
This is Waymo’s 3rd “involved” fatality. While those who study crash statistics generally count “involvement” rather than fault, it is necessary to look at fault with robocars because they are not humans where it is difficult to measure fault and not necessary.
Superhuman Crash Prevention
Today, it would be foolish to demand that early robocars be held to a much higher standard than human drivers when it comes to fault. However, as the field advances, we should be interested if there were things a robocar could do, with its superhuman sensing or computation, to prevent crashes humans would not have prevented.
Waymo studied this, and in 2021 released the results of a simulation of all known fatal crashes in their operation area in Chandler, AZ. Those crashes involved human drivers of course, usually in at least 2 cars. Waymo simulated what would have happen if their Waymo Driver had been in control of the at-fault vehicle, but also if it has been driving the other vehicles. They found that they eliminated all the crashes if they played the at-fault driver, and 82% of those when they were driving the cars that got hit. The former is straightforward: The Waymos did not make the same mistakes these humans made. The latter indicated they drove more defensively and also actively avoided being hit. For example, in many scenarios the at-fault car would run a red light, and the simulated Waymo would see this and hold back going forward on the green, and not get hit.
If we look at several significant incidents, including this one, the question can be asked if the vehicle could have prevented or mitigated the crash. Sometimes this is apparent only in hindsight. Other times it seems good general practices might save the day.
This Dallas Fatality
While we don’t have full detail on the Dallas fatality, we note that the Waymo began braking when it concluded there was trouble, and as such the impact with the pedestrian was low speed. In theory, with a good physics model, the car might be able to conclude it should not brake, or should even speed up. After all, if the pedestrian hit the back side of the Waymo after it had slowed to 5mph, it seems possible that had the car been going faster (not braking at all or even accelerating with electric car speed) it would have been well past the scene before the pedestrian arrived. There was other traffic in the other lane, so the Waymo did not have the option to swerve, but otherwise that might happen.
It should be noted that it’s possible that there wasn’t enough information to determine that reliably, and that braking is generally the wisest course and speeding up is risky and would require strong certainty it was the right thing to do. But what’s interesting is that there are situations (not necessarily applicable here) where a car with superhuman physics abilities could take steps a human could not.
For example, Waymo first noticed the crossing 150 feet ahead. At 24mph a car in full braking a car can stop in just 28 feet from when it applies the brakes, and there were no passengers. However, there were vehicles beside and behind the Waymo which could make a sudden stop risky (even though legal.) Even so, if the vehicle was able to predict serious danger in advance, it could have stopped. The reality, though, is that pedestrians cross streets outside of crosswalks all the time. In Texas (and all other states) the car has the right-of-way in that situation, though a driver mast practice “due care” and do what they can to avoid hitting such pedestrians. The result is a sort of dance between cars and such pedestrians, but it is not expected cars will simply stop because a pedestrian has entered the other side of the roadway, or that they be ready for a situation like this where the pedestrian was hit and thrown.
Because the impact was very sudden, it appears the Waymo had very little time and little information about what was about to happen. The superhuman computer might have been able to predict the impact it it were aware the pedestrian had his back to the SUV, but that’s not an easy thing to do, particularly at night. (It’s quite common for pedestrians to step on to the street and then retreat if they see a car advancing, but pedestrians sometimes make serious mistakes, as this one may have.)
As such, my initial conclusion is there’s not much, if anything, the Waymo could have reliably done to prevent this Waymo-pedestrian contact, unless it was particularly good at reliably predicting the first SUV impact well in advance. Once the pedestrian was flung into the Waymo’s lane, a decision to speed up would not have been safe, or even effective.
Waymo will probably add this situation to their simulation scenarios, and see if there was anything they could have done, or any other way to improve such situations. Waymo has not yet done this, but it’s just been 2 days. Every incident with robocars is a chance for improvement, though this a tragic one.
Cruise Dragging
Cruise was not at fault for the initial impact. Fault lay with the pedestrian crossing against the light, and more with the hit-and-run human driver. But the Cruise could possibly have noticed the pedestrian in the crosswalk, who was acting in an erratic manner. She crossed the Cruise’s lane and the Nissan’s lane and tried to enter the oncoming lanes but reversed direction in fear of oncoming traffic, according to reports. This will happen during a “jaywalking.” While cars can’t brake every time a pedestrian tries to cross the road against the rules, it would have been wise for the system to identify that something dangerous was going on ahead. It did not need to accelerate quickly from the line on the green light. This is particularly true since the vehicle was vacant and there was nobody behind it at the light. It could be more cautious in that situation than otherwise.
Had it done so, the Nissan would still have struck the pedestrian and thrown her into the Cruise’s lane, but the Cruise would have been able to stop before hitting her, and of course would not have dragged her. It was never revealed how much of her injuries were do to the Nissan, the Cruise, and then the dragging. Her name and details were not revealed, and Cruise made a settlement which was reportedly large.
Waymo Motorcycle Rear Crash
A Waymo was slowing to turn right into a driveway. Due to pedestrians in the driveway, it had to slow much more. The motorcyclist behind did not stop in time–officially following too closely–and hit the Waymo, then was thrown into the next lane, where a hit-and-run driver (again!) struck and killed him. Again, no legal fault to the Waymo
At the same time, it has an obvious option here. When it sees the motorcycle behind is following too closely and not stopping, it can just abort its turn into the driveway and speed up with EV speed. Quite probably the whole incident doesn’t happen, and the Waymo has to loop around to get into that parking lot to get its passengers. Humans aren’t constantly looking behind them and predicting the actions of road users behind them, but robot are.
It’s notable in all these crashes that the vehicles were vacant. That creates an opportunity and a problem. A vacant vehicle is free to hard brake, hard accelerate and hard swerve with no fear of disrupting or hurting passengers. A vehicle without other cars around it also has extra freedoms. It raises questions, however, when a vehicle can’t take actions because of passengers or other situations. It’s always been possible to have vehicles drive much more conservatively when empty and/or with nobody behind them. But the overall result would be unacceptably slow service for passengers waiting for pick-up, for example. And you don’t want to answer a question like, “So why didn’t the vehicle take it slow with that dangerous situation ahead of it” with “we didn’t want to get in the way of the driver behind us.” Drawing a line is difficult, and in court, they only consider the times you drew the line wrong. That could drive developers to keep things simple, and not alter behavior based on these factors.
