Is It Time to Rethink Autonomous Kit?

Alphabet/Google departed the scene years ago, spending how much? Certainly in the billions.

Apple has abandoned the field, after spending $10 billion over 10 years to get nowhere. “But we were so close”.                                                                                                                                                          

Tesla battles on, making tyro mistakes.

After seven years – trying to fit through an aperture for which it was too large – the gap between the tarmac and the belly of a business jet (admittedly parked in a carpark, but on top of previous reports about trying to drive under the belly of a fuel tanker).

After eight years:

·        Not observing that the gate of a gated community was closed.

·        Slamming on the brakes when the traffic light turned yellow, even though the following driver thought there was adequate time to cross (the Tesla owner is assured photos are used in the training of the ANN - photos of the prediction for a continuing process will be of little help in training a static device).

·        Not understanding simple signs, like Entry and Exit.

From someone in the Tesla design team:

A suitable word for this is Foolishness. Attempting to train a static device to handle all the vagaries of driving is a hopeless task, which should have been obvious within a few weeks of starting the project. Not attempting to study the task first is even more damning.        

The only similarity between a real Neural Net and an Artificial Neural Net are the two words in the names.

A real neural net has feedback, feedforward, the ease of assembling a predictor for a new situation in a fraction of a second, many separate processes easily joined together and communicating with each other – visual, aural, threat detection – avoidance  of a collision course, a rear-ender, the grill of a gate. An artificial neural net is a directed resistor network, with all the limitations that suggests. No amount of tinkering will allow it to match a real neural net – filtering static data is more its bailiwick.

Sightlines

It would be great if streets and roads were laid out in rectangular blocks, with sightlines sufficiently long that an intersection can be crossed with confidence. Unfortunately, residents are against having their streets destroyed and rebuilt, Instead, they insist that an AV have a similar ability to a human driver in working out when it is safe to proceed.

This is the merging intersection of the Kingsway and President Avenue at Caringbah (a southern suburb of Sydney). As the intersection occurs between a straight line and a curve, the sightlines are messy. Instead of being able to look for movement down a line of vehicles, you can only see a few vehicles before they turn and disappear behind a line of stationary vehicles. It becomes necessary to draw an inference from indirect information – something that the Unconscious Mind in a human driver does well.

It is understandable that AV manufacturers do not want malevolent drivers to train the vehicle to “suicide” itself, then make a claim for millions – “It could have been us in that car!”. The inability to adapt to new situations (in the way a person so easily does) is a serious impediment to their introduction.

Reading Signs

The owner of the newly upgraded Tesla reports the Tesla driving across the lane with the Entry sign, and attempting to enter the shopping complex through the Exit lane. Signs can start simple, then get complicated.

 

 

Signs can directly address the driver’s Unconscious Mind (see Lane Following)

 

 

 

The sign assumes that the system reading the sign knows colours, knows what a Lane Marking is, understands "use" as a verb, and knows the sign does not apply if there is only one set of lane markings. The system is definitely not constructed from ANNs, but from Active Semantic Structures.

It seems obvious to take this further and have the “Road Rules” for the AV written in English, so the system would have a vocabulary that is dense with concepts and objects related to driving, and not much else (all the common words in English - "and", "the", "on" - are already present for reading the complex text of legislation). “Cyclist” has as definition a person riding a bicycle (which in road terms is a fragile, unstable machine, but very valuable while it supports a rider), with the gap that must be maintained during passing, and values for acceleration, braking, and maximum speed for the flat, uphill, downhill. Similarly for trucks, the jack-knifing of semis, a B-train, the frequent stopping of buses, behaviour in the wet or on icy roads, the screeching of tyres under heavy braking.

If this sounds expensive, remember $10 billion over ten years, and complete failure. If the AV is to compete with, and mingle with, human drivers, it can’t be crude. It also can't copy the behaviour of other vehicles - that might work well while AVs are rare, but would be disastrous when they become commonplace.

P.S.

While writing this, came home by taxi on a dark night - had to turn and cross a busy road to reach a side road. Busy road seemed clear, until a black car without lights loomed out of the dark, visible only by a few gleams from shop lighting across the road (a tree prevented normal street lighting from illuminating it). A vision system tuned to what cars look like would not have detected it, but a general purpose vision system (human eyes) had no problem. Treating something like this as an "edge case" that can be ignored is a good way to get the occupants killed.