Back in December a GM Cruise car had an accident with a lane-splitting motorcyclist in San Francisco. I didn't report on it because the police report blamed the motorcyclist, but the accident is possibly more complex, and the motorist has filed a lawsuit against GM..

That's not too surprising. As the first lawsuit against a self-driving car company, it assures fame for the lawyers, and involves a super deep pocketed defendant keen to come out of it looking good. One could be motivated to sue even if clearly in the wrong.

The accident is a somewhat complex one, which means there is some dispute over whether the biker was in the wrong. He was "lane splitting" -- an action illegal most places but not in California. The Cruise vehicle moved to change lanes to the left. The biker then moved to pass on the right at 17mph, faster than the car going 12mph. The car decides it can't complete the lane change and aborts it. GM says it is "re-centering itself" and the biker glanced the side of it.

It's not clear from this description just what happened. Did the Cruise get all the way into the left lane and then come back to the right lane, getting in the way of the biker? Was the biker on the right side of the right lane, happy to see the car leaving to the right, then surprised when it came back? We don't know from the text description, but the vehicle's LIDAR and camera logs will have recorded everything in great detail.

That makes this part of the question boring. While GM has not published the data, they know exactly what happened, and in court all sides will know. There will be zero doubt on who violated the vehicle code. If it was the biker (who was cited by police as unlawfully passing on the right) things should be over fast.

If it was GM, then things should also be over fast because I can't imagine them wanting to fight that in court. They will make a more than generous settlement. The only thing that should make this thing not be over fast would be if the biker gets ambitious and does not accept even a generous settlement.

Regardless of who is at fault...

While deciding vehicle code fault should be quick and unexciting, there are a few interesting questions to learn:

1. Even if the biker is at fault, is there something about the robocar's behaviour that was different from a human's which nonetheless could be a contributing factor to the accident?
2. Did the robocar fail to perceive the biker, even if they had no duty to? Humans also fail to see bikers all the time, and we want robocars to do better at it. Lane splitting is a specialized issue in California, but everybody wants to drive in California.
3. When did GM release the data to the plaintiff and/or police, or when will they? If not immediately, why?
4. If the plaintiff gets "ambitious" what does that say about other ambitious plaintiffs?
5. In particular, is this accident one which would be resolved quickly and cheaply if humans were involved, but ends up being long and expensive because a robot and megacorporation are involved?

I went into many of these questions in depth in my article on accidents.

Aborted lane change is unusual for a robocar

The safety driver did see the biker, and took control, but too late. That might indicate a failure in perception (not seeing the biker) or a lack of a plan for what to do in this unusual situation. Aborting a lane change is already a highly unusual move for a robocar. Generally they will only initiate lane changes that they are confident can be completed. It could be another driver unexpectedly also attempted to move into the same space. If that driver did not signal, they might hold some fault, though it sometimes happens that two drivers simultaneously both try to change into the same lane from opposite sides. In theory this should not happen if both are signalling and both are looking at signals. The Cruise care will certainly have signalled. What I don't know is how well it detects the signals of other cars to its left. Not all cars have a 360 degree camera view; some only have cameras facing forward.

I presume that sometimes an accident can happen when one car starts a lane change, then needs to abort because somebody else is going for the same space, and then sees somebody on the other side is also already trying to take the space they were vacating. The driver of that 3rd car might be at fault, trying to take a space before it is fully vacated, but it may be there are situations where nobody is at fault because of what computer engineers would call a "race condition."

We'll learn more on that soon, I hope. But let's consider point #1. Several people have suggested that other accidents where the human driver was at fault also fall in this class. For example, there are few instances of human driven cars hitting robocars that are stopped at lights and being more timid than a human would be. The human driver expects the car to go; it is conservative and stays stopped, and it gets hit. Clearly the fault of the rear driver but also a result (not fault) of the robocar driving more conservatively than expected.

Here we might wonder if aborting the lane change was a necessary move, or the action of a highly conservative driving strategy. The software might decide that if it sees any risk to the lane change (perhaps because of car being out of place) that it should do "the safe thing" and return to where it was. That is not the safe thing if an unseen lane splitting bike is moving into that space.

Simulation is the solution

One way to fix this is with simulation. Many teams, and I presume Cruise, have built simulators where they can try out their software in 10,000 variations of this situation. This is exactly the sort of thing you can't test in real life, but if Cruise is doing their job, they should have already tested what their software will do in unusual situations like those that might trigger an abort of a lane change. Or they will be doing that soon -- let's not forget these are still products in development, and they have safety drivers ready to take over for just that reason. The photo above, from a country with much higher motorcycle density, shows how important and complex those simulations must eventually become.

Fortunately, this was low speed and injuries are minor, though the plaintiff is now claiming they are more serious. This will be interesting to watch -- but in theory it should be over very quickly.