Today I’m going to examine how you attain safety in a robocar, and outline a contradiction in the things that went wrong for Uber and their victim. Each thing that went wrong is both important and worthy of discussion, but at the same time unimportant. For almost every thing that went wrong Is something that we want to prevent going wrong, but it’s also something that we must expect will go wrong sometimes, and to plan for it.
In particular, I want to consider how things operate in spite of the fact that people will jaywalk, illegal or not, car systems will suffer failures and safety drivers will sometimes not be looking.
First, an update on developments.
Uber has said it is cooperating fully, but we certainly haven’t heard anything more from them, or from the police. That’s because:
- Police have indicated that the accident has been referred for criminal investigation, and the NTSB is also present.
- The family (only a stepdaughter is identified) have retained counsel, and are demanding charges and considering legal action.
A new story in the New York Times is more damning for Uber. There we learn:
- Uber’s performance has been substandard in Arizona. They are needing an intervention after 13 miles of driving on average. Other top companies like Waymo go many thousands of miles.
- Uber just recently switched to having one safety driver instead of two, though it was still using two in the more difficult test situations. Almost all companies use two safety drivers, though Waymo has operations with just one, and quite famously, zero.
- Uber has safety drivers use a tablet to log interventions and other data, and there are reports of safety drivers doing this while in self-drive mode. Wow.
- Uber’s demos of this car have shown the typical “what the car sees” view to both passengers and the software operator. It shows a good 360 degree view as you would expect from this sort of sensor suite, including good renderings of pedestrians on the sidewalk. Look at this youtube example from two years ago.
- Given that when operating normally, Uber’s system has the expected pedestrian detection, it is probably not the case that Uber’s car just doesn’t handle this fairly basic situation. We need to learn what specifically went wrong.
- If Uber did have a general pedestrian detection failure at night, there should have been near misses long before this impact. Near misses are viewed as catastrophic by most teams, and get immediate attention, to avoid situations like this fatality.
- If equipment such as LIDARs or cameras or radars were to fail, this would generally be obvious to the software system, which should normally cause an immediate alert asking the safety driver to take over. In addition, most designs have some redundancies, so they can still function at some lower level with a failed sensor while they wait for the safety driver — or even attempt to get off the road.
- The NYT report indicates that new CEO Dara Khosrowshahi had been considering cancelling the whole project, and a big demo for him was pending. There was pressure on to make the demo look good.
In time, either in this investigation or via lawsuits, we should see:
- Logs of raw sensor data from the LIDAR, radar and camera arrays, along with tools to help laypeople interpret this data.
- If those logs clearly show the victim well in advance of impact, why did the vehicle not react? What sort of failure was it?
- Why was the safety driver staring at something down and to the right for so long? Why was there only one safety driver on this road at night at 40mph?
That the victim was jaywalking is both important and unimportant
The law seems to be clear that the Uber had the right of way. The victim was tragically unwise to cross there without looking. The vehicle code may find no fault with Uber. In addition, as I will detail later, crosswalk rules exist for a reason, and both human drivers and robocars will treat crosswalks differently from non-crosswalks.
Even so, people will jaywalk, and robocars need to be able to handle that. Nobody can handle somebody leaping quickly off the sidewalk into your lane, but a person crossing 3.5 lanes of open road is something even the most basic cars should be able to handle, and all cars should be able to perceive and stop for a pedestrian standing in their lane on straight non-freeway road. (More on this in a future article.)
The law says this as well. While the car has right of way, the law still puts a duty on the driver to do what they reasonably can to avoid hitting a jaywalker in the middle of the road.
That Uber’s system failed to detect the pedestrian is both important and unimportant
We are of course very concerned as to why the system failed. In particular, this sort of detect-and-stop is a very basic level of operation, expected of even the most simple early prototypes, and certainly of a vehicle from a well funded team that’s logged a million miles.
At the same time, cars must be expected to have failures, even failures as bad as this. In the early days of robocars, even at the best teams, major system failures happened. I’ve been in cars that suddenly tried to drive off the road. It happens, and you have to plan for it. The main fallback is the safety driver, though now that the industry is slightly more mature, it is also possible to use simpler automated systems (like ADAS “forward collision warning” and “lanekeeping” tools) to also guard against major failures.
We’re going to be very hard on Uber, and with justification, for having such a basic failure. “Spot a pedestrian in front of you and stop” have been moving into the “solved problem” category, particularly if you have a high-end LIDAR. But we should not forget there are lots of other things that can, and do go wrong that are far from solved, and we must expect them to happen. These are prototypes. They are on the public roads because we know no other way to make them better, to find and solve these problems.
That the safety driver wasn’t looking is both important and unimportant
She clearly was not doing her job. The accident would have been avoided if she had been vigilant. But we must understand that safety drivers will sometimes look away, and miss things, and make mistakes.
That’s true for all of us when we drive, with our own life and others at stake. Many of us do crazy things like send texts, but even the most diligent are sometimes not paying enough attention for short periods. We adjust controls, we look at passengers, we look behind us and (as we should) check blindspots. Yet the single largest cause of accidents is “not paying attention.” What that really means is that two things went wrong at once — something bad happened while we were looking somewhere else. For us the probability of an accident is highly related to the product of those two probabilities.
The same is true for robocars with safety drivers. The cars will make mistakes. Sometimes the driver will not catch it. When both happen, an accident is possible. If the total probability of that is within the acceptable range (which is to say, the range for good human drivers) then testing is not putting the public at any extraordinary risk.
This means a team should properly have a sense of the capabilities of its car. If it’s needing interventions very frequently, as Uber was reported to, it needs highly reliable safety driving. In most cases, the answer is to have two safety drivers, 2 sets of eyes potentially able to spot problems. Or even 1.3 sets of eyes, because the 2nd operator is, on most teams, including Uber, mostly looking at a screen and only sometimes at the road. Still better than just one pair.
At the same time, since the goal is to get to zero safety drivers, it is not inherently wrong to just have one. There has to be a point where a project graduates to needing only one. Uber’s fault is, possibly, graduating far, far too soon.
To top all this, safety drivers, if the company is not careful, are probably more likely to fatigue and look away from the road than ordinary drivers in their own cars. After all, it is actually safer to do so than it is to do in your own car. Tesla autopilot owners are also notoriously bad at this. Perversely, the lower the intervention rate, the more likely it is people will get tempted. Companies have to combat this.
If you’re a developer trying out some brand new and untrusted software, you safety drive with great care. You keep your hands near the wheel. Your feet near the pedals. Your eyes on the lookout. You don’t do it for very long, and you are “rewarded” by having to do an intervention often enough that you never tire. To consider the extreme view of that, think about driving adaptive cruise control. You still have to steer, so there’s no way you take your eyes off the road even though your feet can probably relax.
Once your system gets to a high level (like Tesla’s autopilot in simple situations or Waymo’s car) you need to find other ways to maintain that vigilance. Some options include gaze-tracking systems that make sure eyes are on the road. I have also suggested that systems routinely simulate a failure, but drifting out of their lane when it is safe to do so, but correcting it before it gets dangerous if for some reason the safety driver does not intervene. A safety driver who is grabbing the wheel 3 times an hour and scored on it is much less likely to miss the one time a week they actually have to grab it for real.
That it didn’t brake at all — that’s just not acceptable
While we don’t have final confirmation, reports suggest the vehicle did not slow at all. Even if study of the accident reveals a valid reason for not detecting the victim 1.4 seconds out (as needed to fully stop) there are just too many different technologies that are all, independently, able to detect her at a shorter distance which should have at least triggered some braking and reduced severity.
They key word is independently. As explained above, failures happen. A proper system is designed to still do the best it can in the event of failures of independent components. Failure of the entire system should be extremely unlikely, because the entire system should not be a monolith. Even if the main perception system of the car fails for some reason (as may have happened here) that should result in alarm bells going off to alert the safety driver, and it should also result in independent safety systems kicking in to fire those alarms or even hit the brakes. The Volvo comes with such a system, but that system is presumably disabled. Where possible, a system like that should be enabled, but used only to beep warnings at the safety driver. There should be a “reptile brain” at the low level of the car which, in the event of complete failure of all high level systems, knows enough to look at raw radar, LIDAR or camera data and sound alarms or trigger braking if the main system can’t.
All the classes of individual failures that happened to Uber could happen to a more sophisticated team in some fashion. In extreme bad luck they could even happen all at once. The system should be designed to make it very unlikely that they won’t all happen at once, and that the probability of that is less than the probability of a human having a crash.
More to come
So much to write here, so in the future look for thoughts on:
- How humans and robocars will act differently when approaching a crosswalk and not approaching one, and why
- More about how many safety drivers you have
- What are reasonable “good practices” that any robocar should have, and what are exceptions to them
- How do we deal with the fact that we probably have to overdrive our sensor range on high speed highways, as humans usually do?
- More about interventions and how often they happen and why they happen
- Does this mean the government should step in to stop bad actors like Uber? Or will the existing law (vehicle code, criminal law and tort law) punish them so severely — possibly with a “death penalty” for their project — that we can feel it’s working?