More thoughts on the self driving car

Recently Tesla had shown a teaser video of their "self driving car" project which immediately drew media attention and swarms of self driving "enthusiasts" to again announce that this is a done deal already (which it is not). Here is the video in question:

Note: above video has been subsequently taken down, I'm now linking to a mirror.

Now this looks very impressive as a demo but there are a few details I'd like to point out before we start saying again that the self driving car is a done deal from technological point of view. Disclaimer: I do like Tesla and I think some of their ideas are great, but their self driving seems a bit premature, somewhat over promised and over hyped.

  • The lighting conditions in a video are perfect from computer vision point of view. Although it is a bit foggy, the illumination is uniform and diffused. There are no hard shadows, flares or ghosts.
  • The lane markings are all clearly painted and visible everywhere.
  • There are no "unusual situations" (see below what I mean by that).

Just a reminder that a self driving car was demoed as a research project in mid 80's at CMU (notice similar weather conditions in the video below):

Back then it was slow (due to the computing speed), but tech-vise it was essentially the same thing as these days. They even used a multilayer perceptron (today it would be called a "deep learning" module) to train the driving behaviour associating visual input with driver commands.

So to summarise:

Things have undoubtably improved since the 80's and Tesla has a great demo. But there are some really major challenges before that demo can become a real product. Aside the obvious challenges such as difficult lighting (hard shadows, flares), weather (wet windshield, distorted radar readout, lane markings invisible due to reflections) there are slightly less obvious but not uncommon, even in very organised countries:

  • miss-painted lane markings
  • vandalised, damaged or covered with snow/ice/mud street signs and signals
  • wild animals crossing the road (often with unpredictable behaviour)
  • items blown into the road by wind such as sand, bushes, plastic bags - some of which can be dangerous to the vehicle, others are benign and it is safer to collide with them rather than to take any sudden action
  • urban environment weird stuff such as intoxicated people behaving erratically, criminals, art performances, protests, riots etc.

Admittedly, many of these factors are confusing to people as well. Humans however can typically collect from many different sources of weak evidence and context to disambiguate and understand the situation. A human (as well as many animals) will perfectly tell apart a benign object floating in the air, such as s leaf or a plastic bag from a brick falling off a truck (based on the forward model of physics). A human will understand the signs indicating construction and a detour and consequently will figure out the combination of old and new lane markings. AI does not have that subtle understanding (as of now at least) and will consequently be confused by many such situations.

Now it is often stated that this is just a minor software problem that will somehow magically fix itself in the next couple of years with the aid of some deep learning sauce, but I dare to doubt it. The approach is wrong, and it will fail like it did many times before. However I do believe that Predictive Vision Model (PVM) would help as it has the potential to learn the physics and natural behaviour of objects, but admittedly it is still "researchy" and may need a few years of research and scaling up before it can show applications in the automotive domain .

For all those remaining "self driving car" hypesters (I would actually call myself an enthusiasts) here is a summary how the self driving beer delivery in Colorado was organised:

Notable quote for those lazy:

Our protocol was that the truck had to make this trip eight times with a driver [in the cab] without the driver taking over before we would allow the computer to drive the whole route [with the driver in the sleeper compartment].

Our involvement was multifaceted. We made sure the roadway was swept of debris, and we got our tow trucks involved to make sure there weren’t any abandoned vehicles on the side of the road that would hamper the test. We worked with our road construction contract partners to make sure there weren’t any unexpected road projects going on that would cause weird traffic conditions. The highway patrol had a lot of questions about how they would pull over a truck with no driver in the seat and how we communicate effectively with the autonomous vehicle.

Makes one wonder if Tesla demo was indeed "organic" or was it "organised" (staged) in a similar way. I personally believe it was organic, and therefore is very impressive, but there is still a long way to go.

Update (Jan 2018): It appears that the above Tesla demo was just as fake as the Otto truck. The discussion here sheds more light in the face of California DMV disengagement report.

If you found an error, highlight it and press Shift + Enter or click here to inform us.