Autonomous Vehicles 2021 : How do we get there - fast?

 First item - analysis paralysis:

Engineers will pile on so many what-ifs that the problem becomes impossible. Try to do too many things in the kitchen at once and everything burns. So how to we get past that? We build what we have and see how it plays.

Bottom Line - Platooning:

Start with driver in front car similar to Uber driver, the 4 cars behind are platooning, like pack animals following in a line. Preroll recording, maybe 15 seconds to one minute in advance and when event occurs, save that 15 seconds until the end of the event.

 

This may not be conveyed perfectly since I am thinking out loud. 

What I am attempting to cover is how to optimally collect data for control of autonomous vehicles based on current drivers. This is necessary to allow the transition from the human driver, to humans and machine drivers, potentially to entirely machine drivers.

The first premise is that there are an infinite number of variations or combinations of outcomes leading up to an event and possibilities following an event. This can be compared to nature where there are millions of combinations possible (Raup’s morphospace), but few of these occur in nature. For data collection it is best that we start with those that exist. In this space, we do this by capturing the data intensely around a given event. 

Capturing all of the data all of the time, which we can do, leads to more analysis, generating what we see in the military as a self-licking ice cream cone. It is better to collect data around a given event, by pre-recording in a loop 15 seconds to a minute, and when an event occurs, saving that data and on through the end of the event - intensely. Then the data is not a continuous video, but a set of short clips or events that can be cataloged. These clips or vignettes  are easily understandable, are based on actual experience, and serve as the core of the actions/reactions for the future autonomous cars.

Where to start?: At this point, platooning is a real possibility. A lead car is driven by a human (a professional driver) and the other cars (we suggest 4 max at this point) follow behind the lead - autonomously.  The follow cars take their guidance from the moves of the lead car, which should experience events slightly before the trail cars. This greatly simplifies the initial development of autonomous driving, while giving the trail car driver and passengers the experience of full autonomous driving while on the highway or in traffic. (There is a great deal of variable, “unpredictable” behavior on the road that a professional driver will negotiate as a matter of course and instinct, which a new autonomous system will struggle with.)

Recording: Record your normal data that you have likely been collecting before. I won’t imagine to list all of this here. Important to add though are the reactions made by the trail cars and the reactions, including physiological, of the lead driver and the passengers in the follow cars. This will build toward user experience.

Why 5 cars?: In efficient movement in traffic, this is about the extent of any type of small convoy. This allows for several maneuvers. When taking left turn in the U.S in formation, the lead car turns wide, the two car turns inside of one, three goes directly behind one, 4 goes inside of three, and five goes inside of the other 4 almost facing oncoming traffic. This maneuver stacks the cars up so that they fill a very narrow space and are able to clear the intersection as one unit. Another move is changing lanes in a peel. All 5 cars are in line in one lane with the one car in front followed by two, three, etc. in sequence. In this case the two car changes over one lane and slows down. The other cars continue, passing the two car in the adjacent lane. Once the five car passes two, one-three-four-five move over into the empty space created by the two car slowing and the two car becomes the 5 car. 

One more piece of efficiency comes from pulling out of lights. Right now the cars each start in turn from the front. There is a delay associated with each start. This leads to a slinky effect. The same happens on slowing as experienced day in and day out in current traffic. In highly efficient military convoys a call goes out on the radio so that all cars start moving at once, similar to soldiers in formation. In marching formation this starts with a mark time march in place followed by a forward march - getting thousands of troops moving and not falling all over each other or leading to a day long slinky before the last person starts walking.  It is a process that scales, which it will with cooperative autonomous cars eventually. 

There are so many variations in current driving between cities and rural areas and even off-road. Trying to capture that as one huge sample set seems prohibitively expensive and possibly self-defeating in the short term. Recording the actions of active drivers is a good place to start and with platooning this can be done semi-autonomously, providing a safe transition alternative while providing the autonomous experience to the trail car drivers and riders. This is a pattern I have experienced first hand riding on pack animals, which people have used for millennia is this capacity. 

[One added part is a remote safety driver. When autonomous driving does start, there will need to be a remote person (yes there is a communication system delay) who is alert and ready to jump in as the “driver” albeit remotely to control the car for 18 seconds to 2 minutes until the actual driver can get their bearings. This comes from the idea in combat, or waking a sleeping driver, where in the first 18 seconds the driver is disoriented and almost in shock. Once they have gotten their bearings they are ready to control the car, but in the interim they are a danger to themselves and others by their impulsive reactions.]

I hope some of this is food for thought.