It's a two lane road, and an object in the other lane as a car is driving along isn't unusual. I'm sure the LIDAR saw it but how is the car supposed to know the object is about to move in front of an approaching car? It makes no sense. Somewhere in the software there has to be some assumptions that other road users simply don't do that. The car would be constantly braking otherwise and could never overtake anything.
The real question is: How could the pedestrian not see the approaching car? They just walked straight across the road without even looking. No human would have expected that either.
This... This, this, this, this!
This isn't about if the Lidar was working, it is about what predictions were made based on the available data from all systems and the information about the surroundings.
We ALL make predictions about the world around us based on the rules of reality. How many times would you have been in a terrible accident and likely killed on your way into work this morning if the car next to you on the highway went crazy (broke the rules) and turn hard into the side of your car at 65MPH? Likely 100% of the time. What extra sensors would save you? How many times were you actually in that possible situation in just this one drive to work? 100s of times if not more, in just this one ride? With something that potentially deadly, why is that not the greatest ever present fear in all of us all the time every time we get behind the wheel??? Because we make the prediction that people will stay in their lane and follow the rules of the road. Is it a good prediction? Well at least most of us reading this made it where we were going today (and the vast majority of days like it), but some small amount of people actually did get hit/killed in that same time in this exact way. This is the fault of the person that broke the rules by which we make predictions, not the person expecting the rules to be followed. Just because it's a computer and not a person, and just because they have more sensors, doesn't mean they aren't making the exact same predictions, and relying on everything around them to fundamentally follow the rules.
Now in this particular case you need to look at the situation and the reasonable predictions being made.
The car does have an impressive set of sensors, and under optimal conditions those sensors all contribute to the total perception of the surroundings. When multiple sensors contribute the same feedback [camera sees and identifies a person, lidar sees an object of some sort (lower rez), radar sees an object (again low rez)] then confidence can be high what some thing is and how you can expect it to act.
Based on the video, the visual light cameras could not see the person in the shadows, at least not well. Mark that off as one of the sensor contributions to the feedback. The camera based computer vision is probably the most capable of determining what something actually is since it has the highest resolution representation of the details of a thing. Colors, faces, textures, etc are all the domain of the computer vision that the other sensors can't "see".
The camera system, even multiple cameras in an array, is less good at distance and velocity determination. That's where the Lidar comes in. Excellent quick distance measurement but terrible resolution. It gives you a point cloud, which some computer has to put back together and.. .you guessed it... make a prediction of what that object could be. Did the Lidar detect there was some sort of object over there? We don't know yet, but I would guess it probably did. The question now becomes, without the visual camera data what did the Lidar see and what assumptions did the computer make based on that sensor data about the probable object?
Question1: What can be reasonably expected to be in the left lane of a two lane highway in the middle of the highway far between street crossings?
Answer1: If you said a person with no regard for their own surroundings wandering between lanes at night dragging a bike loaded up with stuff with no lights on, then you are crazy. That is a poor prediction by person or computer. Probability of that is low. Very low. It would be a better prediction to think it was a couch or some other junk that fell off someone's truck that's blowing around in the street, which is more probable.
A much better prediction would be a motorcycle of some sort. Probability of that is fairly high. You would expect a motorcycle to be in the street, and expect it to be in the left lane, possibly stopped or moving slowly in the beginning of a left turn lane. In the place where the accident took place, the highway opened up from 2 lanes into 4 lanes with 2 left turn lanes on the left.
Question2: What actions can you reasonably expect a motorcycle to take while in a left turn lane?
Answer2: If you said purposefully swerve into a car driving at a normal highway speeds in the right lane without any regard or hesitation, then again, you are crazy. Probability is incredibly low. A much better prediction would be merging back into the left through lane from the beginning of the left turn lane. Probably of that is good. This would give you the profile of a side view of a motorcycle moving at slow speeds, which is something the Lidar could tell on it's own and matches what we see in the video.
Now I'm not saying this is a perfect exoneration of the computer. For example; is it a good prediction for a motorcycle to have no lights on at night? Probably not, but it's also not a terrible prediction. Vehicles are pretty routinely found driving at night with no lights on.
Should the computer have realized it has sub-optimal sensor information without the full camera detail, and there was a questionable condition of a vehicle with no lights on in the adjacent lane, and therefore proceeded with more caution? Maybe, but not 100% certainly does that mean it needs to take evasive action. Maybe it did reduce it's speed to what it considered optimal based on the predictions in the seconds before the video. I'm sure we will hear from the Uber team eventually, but if the prediction was a vehicle such as a motorcycle (which I think is a reasonable prediction by human or computer standards given the conditions and probabilities) that doesn't warrant the follow up prediction people are making here that the object will try to jump in-front of the car and demands panic breaking. Now why the breaks didn't lock up once it was clear the object was going to get hit is another question, but it might come out that they did and it's just not obvious from the video. I'm willing to give them the benefit of the doubt on that one since it's not clear. Dave was the one that synced the driver with the front video and that could be off.
Here is a completely incomplete list of things we (or a computer) would never attempt if we had to be 100% certain of every prediction or be able to take 100% successful evasive action:
1) Never enter a building of any kind. It could collapse, it could light on fire, an airplane could crash into it, etc.
2) Never turn on a light switch. You could be electrocuted, you could start a fire, etc...
3) Never walk on the sidewalk. A car could hop the curb, you could fall in a hole, a tree could fall on you, etc...
4) Never drive on a highway. 65MPH is too fast to react to all possible situations.
5) Never drive on a city street. 35MPH is too fast to react to all possible situations.
6) Never get in your car at all. Breaks could fail, steering wheel could pop off, gas could explode, etc....
All of these things WILL happen, but that doesn't mean we need to account for them in everyday predictions. Self driving cars will kill people. Period. It will happen fairly often when the technology is more prevalent. There is no way around that with 2ton steel boxes moving at high speeds. The question is whether they kill people less than people drivers kill people, which will most certainly be the case.
TL;DR
Don't break the rules and walk in-front of moving cars. More sensors doesn't ensure perfect predictions 100% of the time.