Most cars have some Level 1 automation, such as cruise control and various skid/flip probability reduction systems. Most high volume passenger vehicles have higher levels. Some military and private air, land, and sea equipment are already at Level 5.
Level 4 requires that driving be automated during normal driving conditions, with manual override. However, to my knowledge, no one has published a mathematically terse and comprehensive distinction between normal and abnormal to aid in testing Levels 4 vehicles, so testing to a standard is probably not yet possible.
For legal and political reasons, Level 5 is essentially a statistical criteria. For fully automatic to be viable as a market product for general public use, the safety data for passengers and pedestrians must indicate at least the level of safety for manually driven vehicles. Although this will likely suffice from a law and public relations purpose, it is inadequate as a quality standard for automation engineering and testing. The ambiguities are numerous.
- Statistical criteria required to pass the test (i.e. sample size, duration, randomization, and double or single blindness)
- Mathematically terse and comprehensive scenarios for the test
- Allowable proportion of level 2, 3, or 4 vehicles in the control group.
- Probably others
There will be no vehicle driver in what they are calling Level 5 — only passengers. The idea is to give no power to the occupants of the AV other than destination or change in destination.
This has been the safety norm in other sub-sectors of transportation for a century. Passengers cannot talk to the pilot of a jet or the engineer of a train. In the majority of cases, safety is compromised whenever a person that has not undergone the discipline of intense safety training has control over any aspect of the vehicle's operation.
That is the primary impetus behind AV from the forward thinkers in government and academia.
Specificity and Insight
It is of paramount importance that researchers define system criteria more specifically and scientifically. Systems architecture, software engineering, safety evaluation, and quality control policies and procedures of the automated system driving the AV requires such.
With a billion plus lives at stake, the design should progress with the diligence and care as if designing a human-occupied drone aircraft or a civilian Mars lander from the ground up, even if your first phase is to only achieve what is being called Level 4 for basic passenger cars.
Target Reliability and Safety
Humans eat food, have emotional conversations, text each other, ingest mind altering substances, and fall asleep while driving. Judging the safety statistics of an AV by comparing with those of humans driving may sound practical, but it is absurd. It will become clear just how absurd the popularized Level 5 criteria is as the parameters of design are enumerated.
Design should instead minimize the possibility of any accident ever. The goal should be zero mistakes, both at point of sale (the dealership) and at later points in the product life-cycle as the AV learns.
Defining a Mistake
A mistake should be defined as follows:
Any less than optimal state indicated by the correction signal used to direct reinforced learning in any of the system's re-entrant or coincident training mechanisms
The adaptive (i.e. machine learning) portion of the system must permit re-entrant or coincident training (reinforced learning) because there is no possible way to predict the common routes of the buyer at the vehicle's point of sale.
To comprehend the complexity of the problem space for AVs and begin to simplify it, consider the dimension of conditions, controls, and priorities (embodied in feedback signals) related to driving on roads for any vehicles that use roads.
- Starters (there are two in the case of most hybrids)
- Engine stop
- Breaking controls (there are two in the case of regenerative breaking)
- Steering shaft or hydraulic control position
- Break control position (three or four depending on emergency/parking break design)
- Transmission planetary gear clutch state, traditional automatic transmission control state, or clutch and traditional manual transmission control state
- External lighting switch positions (Widely variable, but at least six for headlight, high beam, break light, left signal, right signal, and tail light)
- The content of any messages broadcasted, multicasted, or sent specifically to any other vehicles with compatible reception (if birds do it, so can electronic systems) either via light, sound, or RF (this will require the development of layers of inter-vehicle communications protocols)
- Probably others
Data Acquisition Channels
- Wheel positions (there may be 2 or 4 positions to read with an encoder since one cannot assume perfect alignment)
- Break torques (4 of them, which can be read by 16 redundant strain gauges)
- Break metal temperatures (4)
- Torques and temperatures for any independent emergency breaks
- Accelerometers (two devices x three dimensions per device to detect acceleration/deceleration, centripetal force, and, with some math, tire skid velocity for all four tires in two dimensions)
- Tachometers (one before the transmission and one on each wheel)
- Engine and coolant temperature detectors
- Cameras (must be high resolution to recognize animals, humans, shopping carts, curbs, road signs, train signals, speed bumps/humps, and hazards, which can be IR or visible and the more angles covered the better)
- Wind turbulence resistant external microphones to detect horns and sirens (at least four to detect likely orientation of audio source)
- Suspension strain gauges (to detect vertical road force on each tire)
- The content of any incoming messages from compatible systems
- Battery voltages and currents (two batteries for regenerative breaking or hybrid startup and motive assist, and possibly several currents)
- Fluid levels, pressures, viscosity, and transparency (fuel, oil, steering, transmission, break hydraulic, and possibly others)
- Probably others
A system can be operational and possibly reach what most would consider Level 5 with fewer channels of acquisition and control than above, but it would be poor technology planning to start designing systems with unnecessary limitations. Such limitations will also very likely increase the cost of engineering and the effectiveness of training while saving nothing.
Why Vehicles Have Very Little Instrumentation Today
A human cannot make use of all of all the information above. Nor could a human control all the channels listed above without a high frequency of mistakes. A properly designed electromechanical learning system can.
It would be lazy systems architecture not to capitalize on the positive impact the additional instrumentation would have on safety, total cost of ownership, and other quality criteria for the AV buyer who can afford the extra sensors and computing power. Furthermore, after a few years of manufacturing for a mass market, the cost of the additional components will become small in comparison with the cost of metal and plastic.
Operational Criteria and the Formalization of a Mistake
The problem space contains at least nineteen dependent variables (output channels) and forty-six independent variables (input channels). Some are binary, some are floating point, some are streamed data, some are streamed audio, and some are streamed video.
Together they form a space in sixty-five dimensions. That is what must be optimized according to some predetermined and possibly re-programmable formalization of what is optimal.
Let's consider this idea of optimum safety, thrift, and comfort as quality control criteria. Real time quality control should follow TQM ideals, continuously ensuring quality in multiple dimensions and at multiple points in the total system.
- Maximal distance from other vehicles
- Maximal distance from stationary objects (bridge abutments, buildings)
- Maximal distance from pedestrians on foot, bike, wheelchair, ...
- Maximal distance from edges of pavement
- Minimal lane switching
- Minimal loads and torques on wheels
- Minimal fuel consumption
- Within operation parameters for tires, breaks, engine RPM, and dozens of other parts and subsystems
- Shortest distance for route to destination
- Shortest time distribution mean for route to destination
- Safest route to destination
- Least stops on route to destination
- Probably others
Most of this must be balanced, so optimization criteria must be aggregated. Such aggregation must go beyond the simplicity of a loss function. Summing squares will not work at all, so give up aggregating in such a simplistic way. A multivariate extension of the Pythagorean Theorem is find for calculating distance in linear space, but driving cars is very non-linear. This kind of robotics system will require more thought in the formulation of balances, priorities, and the concept of emergency.
Further expanding on the above definition of a mistake, any real time control that does not optimize the sixty-five-plus dimensional surface is faulty. Now what is optimal must be defined. Consider the following quality control criteria, roughly in correct order.
- Pedestrian safety
- Passenger safety
- Mechanical system integrity
- Fuel conservation
- Vehicle external coating integrity
- Mechanical system wear
- Passenger comfort
- Time conservation in reaching destinations
- Probably others
Applying Optimization in This Context
Aggregation of the incoming signals acquired are not only based on multiple criteria but the prioritization is not always constant, implying the need for a vector of correction signals rather than a single floating point number.
A single dimension of signaling to feed the disparity between ideal operation and the current system behavior (called a loss function in gradient descent) will not suffice. There will, out of necessity, be a need for training and reinforcement with a complexity that involves the idea of preemption. Evolution has declared preemption the design of choice for nervous systems with brains.
For instance, the pedestrian safety feedback signaling must always preempt the fuel conservation feedback no matter how much fuel would be consumed in staying clear from pedestrians, in planning a route where pedestrian density is lower, steering the vehicle clear of people, choosing speed, and applying breaking.
All biological systems have these preemption mechanisms — even bacteria. A turtle doesn't balance the transportation aspect with safety when it retracts under its shell. The behavioral interest in the turtle's destination is shelved (stored and temporarily forgotten) until the preemptive system that detected danger indicates the danger has passed.
Humans Should be Passengers on the Streets of the Future
The reason humans are generally driving in a continuous state of mistaken control is because the priorities that maximize the parameters of transportation for society (above) are inconsistently followed by humans. Birds fly smarter than humans drive. The priorities of an emotional being that wants to get somewhere fast and while talking, texting, eating, and possibly getting high will often be mistaken.
Future people may look back at the period between the advent of Model-T market penetration and the complete transition to AV as a period of strange inequality. Stepping back, the worldwide interest in domestic security, airline safety, train and subway safety, and building code contrasts strongly against the cultural insistence of every household have instant access to drive anywhere, any time, and in any mental condition.