Zoox shows an hour live driving video, impressive and unimpressive


Zoox is the $1B funded startup trying to build a radical design self-driving car. Last week they released a video of an hour long drive through Las Vegas, going through pick-up zones in hotels and the airport. The car does a number of impressive things, but at the same time, showing these things and an hour of driving are only the tip of the iceberg of what you need to do, making the video unimpressive at the same time.

In this new article, I go into what goes on in the video, and what it means.

Read Zoox shows an hour live driving video, impressive and unimpressive


It's not enough to drive as well as the average human. We need these cars to drive at least as well as the median human.

I would not disagree, but I have not seen a great deal of research into data on the median human record. Do you have any figures?

If I had to guess, I would suspect the median human has 1 accident in a lifetime perhaps 2. It is of course a very discrete number. A lifetime of driving is around 500,000 miles, so in fact depending on the type of accident we're talking about, it's not too far off the average. The average human has a police reportable accident every 500,000 miles, but a general accident every 250K and a small ding every 100K, according to study of US drivers.

No direct figures, but 28% of vehicle fatalities are caused by drunk drivers, and I don’t think the median driver ever drives drunk, while the average driver would of course drive drunk some percentage of the time. Add in other types of reckless driving that are done by a minority of drivers and cause a disproportionate number of crashes, and it seems clear that the median driver is much safer than the average driver.

On the flip side, I’m not sure that the best drivers are all that much better than the median.

In any case, I don’t think this is something that can or should be measured. It’s just a theoretical goal.

Interesting review that picks on a lot of non-obvious challenges. Another one of those is night performance. The camera is probably the leading sensor in autonomous perception, but at night they are more or less limited to the coverage area of the headlights (besides detecting other car's tail-lights). Has any AD developer shared significant night footage?

Well, urban driving actually will have most of your main driving areas lit with streetlights, but not all of them. HDR cameras are becoming the norm to be sure you can perceive well at night. The nice thing is the other cars are super obvious at night. The pedestrians are another story. LIDAR of course is great a night, slightly better than it is in the day.

Note as well that when obstacles are lit by your own headlights, the illumination doesn't vary nearly as much as objects in the day which have the sun in all sorts of places, and shadows etc. it's easier to classify something in your headlights -- but only to the distance limit of those lights.

Add new comment