Auto-pilot and autonomous (self-driving) vehicles are poised to become
the defacto technology of the next decade, but not before working through some
considerable technological challenges which, if unaddressed, could eventually
delay or derail the entire venture.
As of this writing, these vehicles, in their current version and state
of development and premature deployment, constitute a serious threat to life
and property, merely because the state of the technology has not yet warranted
or earned status required to operate in live traffic, thereby causing them to
be both deadly and dangerous.
Before any more tragic accidents occur, the vehicles must be pulled
from the live traffic and entered into a long stage of beta development on
closed roads and training facilities, with specially trained human drivers as
Instructors or coaches. These human mentors, teamed with the most sophisticated
A.I. technology, will together rewrite the code that powers the decision-making
capabilities that power these autonomous systems. Further, the cars should not be returned to
the road until all the current bugs have been expunged from their systems,
making them safe to operate around human drivers and pedestrians.
Until this is accomplished, a self-driving car is nothing more than a
novelty, and nothing less than a lethal weapon.
Final Caveat: Until they are
perfectly safe, self-driving, autonomous, or auto-piloted vehicles should sport
warning lights (similar to police or taxicab lights) and issue steady, periodic
warning sounds alerting all pedestrians that such vehicles are under autonomous
control, and potentially dangerous.
Accidents
Involving Self-Driving Vehicles
Incident: Vehicle plowed into a concrete highway lane
divider, and then burst into flames.
Maker/Operator: Tesla
Failure/Cause:
Vision failure with some stationary objects, also objects with color that
blends into background
Fatalities/Casualties: One, driver Wei Huang
Incident: Autopilot slammed into stopped fire truck
Maker/Operator: Tesla
Failure/Cause: System cannot detect non-moving objects
while autonomous vehicle is traveling more than 50 mph/80 kph and the vehicle
ahead moves aside, to be replaced by a non-moving vehicle in the front
position. (Such systems as autopilots are designed to ignore static, non-moving
objects.)
Fatalities/Casualties: None
Incident: Self-driving vehicle kills pedestrian
Maker/Operator: Uber
Failure/Cause: Software bug, pedestrian was detected by
sensors, software ‘decided’ to ignore her, determining her to be a ‘false
positive’.
Fatalities/Casualties: One, a pedestrian was struck as she crossed
the street in front of the vehicle, which did not slow upon approaching her.
Incident: Self-driving car aborts a lane change
suddenly, knocks over motorcyclist moving into previously vacated spot.
Maker/Operator: GM (General Motors)
Failure/Cause: Autopilot ‘misjudged’ when it was safe to
switch lanes.
Fatalities/Casualties: One, motorcyclist knocked from his bike by
car abruptly switching back to previously vacated lane.
Note: In the
case involving the fatality in Arizona, where the car ran into the pedestrian
without slowing (regarding her as a ‘false positive’ to be ignored) “The
National Highway Traffic Safety Administration concluded
that the system was operating as intended, wasn’t defective, and that
Tesla didn’t need to recall any cars.” Although it is true that the vehicle’s
systems operated as intended, it still killed a woman, and, so is still in error, and should
have been removed from the road, including all models with the same driving
systems.
1. Self-driving cars are potentially deadly weapons
and should be treated like as such.
2. Default to Halt: During the beta stages of development (which
should be on the number of years not
months or weeks) certain safety protocols should be in place which immediately
halt the vehicle (as soon as it is safely possible to do so) until it has been
determined (by a human driver or coach)
that it is safe to proceed.
3. Human Instructors, specially licensed and trained human drivers
need to be present in the vehicle throughout the beta stage to make decisions,
keep notes, and pass this information on to the programmers and hardware
developers for continuous and safe evolution
and development of the car’s navigation systems.
4. Specialized
training: Instructors
should also receive regular training along with the vehicles, all with an eye
toward improving the car’s autonomous driving ability by uploading or
‘imprinting’ the decision-making skills of the coach into the car’s A.I.
systems, and having those systems analyze the driving practices of all coaches
and instructors to look for aggregate, repeatable practices to emulate.
5. Extended Beta Stage: While in the beta stage of development,
autonomous vehicles should be designed and built with systems that prohibit the
car from starting, moving, or even turning on unless there is a coach in the passenger seat, logged in to the
car’s navigation system, with his/her own steering and controlling system.
6. Driving and
Training Range: Self-driving cars
should also be routinely subjected to a driving range especially designed for
them, to shake out unforeseen bugs in the system without endangering human lives or property. This debugging stage should not come
at the cost of any more human lives.
7. License Revoked: With pedestrian safety in mind, all autonomous
vehicles now on the road should be suspended from driving in ‘live traffic’
until the end of the beta period.
8. Failsafe &
Secure: The proposed systems, as previously
defined in the initial document, Improved System
for Self-Driving or Autonomous Cars, have several layers of built-in redundancy. At least three (or more)
of these different subsystems (out of the dozen or so) must agree on a course of action before the car proceeds. If these
criteria are not met, then the vehicle should stop, instead of plowing ahead. The
traffic situation, which causes the vehicle to halt, should be fully detailed
by the car’s own systems, and by the on-board, human coach. This information
should in turn be added to a comprehensive database, which will become part of
the ‘experience’ set of all autonomous vehicles. As such, such a database
should be monitored and updated regularly. Eventually, the A.I. learning
systems will parse the data submitted by all
drivers and vehicles, using that data to ultimately learn repeatable, ‘best
driving practices’, which can in turn be passed on to the vehicles’
programming.
9. Mimic Human
Behavior: In time,
autonomous cars can become better drivers than their human counterparts, but
until that time, those vehicles can still ‘learn’ from human drivers, even the
worst of them. Human drivers, on encountering something ‘unknown’ in the road,
will come to a sudden stop, if only to preserve their own lives and property. Autonomous
or robotic vehicles should mimic this behavior, especially since human lives, property, and expensive prototype systems (the car
itself) may be endangered or compromised.
10. Self-Diagnostics:
On startup (with coach present), the car
should run through a series of self-diagnostics for all its navigation and
communication systems. If any of these systems fails, and second diagnostic
should be attempted before the car reverts itself to non-autonomous mode,
secures it doors and windows, and, if it can, call in its status to the nearest
Control Center. During this time, only a Coach or a systems developer should be
able to unlock the vehicle and drive it manually, otherwise, it should be towed
to the nearest certified service facility for a full systems analysis.
11. The Obstacle
Driving Course Police and other
armed safety and security officers often have to train on a specially designed
course called ‘Shoot Don’t
Shoot’ where they are
presented with several ‘targets’ which they must decide, at a moment’s notice,
whether to shoot or not shoot. Such a course, modified to present various
traffic situations, needs to be developed for the A.I. systems, which control
autonomous vehicles. This exercise would allow the software and hardware
developers (and the A.I. itself) to gain useful practice and insight into which
conditions on the ‘real road’ require the vehicle to stop, or not without
endangering human lives. When there is even the slightest doubt, the vehicle should be trained or programmed to
stop.
12. Hardware Sensors are the car’s eyes and ears, and, as such, must
be near flawless. Because no hardware systems are (flawless), redundancy has been prescribed for these systems,
as described in the parent document (Improved System for Self-Driving or
Autonomous Cars). These vehicles
should not be working as human
drivers do, peering ahead into darkness (or foul weather) then reacting too
slowly, too late, or worse, not at all, when something unexpected appears or
occurs on the road ahead of them. The purpose
of having self-driving or autonomous vehicles is to actually improve on human drivers, not emulate their
bad habits. If this is not the case, then why have them? These vehicles
should have and area-awareness far superior to that of humans beings. They
should also react faster, and have access to a constant data stream of road and
weather conditions, which would preclude
them from ever being ‘surprised’. For this to be possible, they must have superior detection systems
with multi-layered redundancy, and a ‘voting system’, where a significant
majority of its systems decides. Any less, and the vehicle stops; it does not
‘plow ahead’; a human driver would not, unless he or she was somehow
incapacitated.
13. Digital Rear View Mirrors should be able to capture and store video to the
diagnostic system or ‘black box’, and also to aid the car or its
owner/Instructor in operating the vehicle should the need arise.
14. Laser -enhanced Headlights can be used to measure the distance between one
vehicle and another, and as an additional
way to detect other obstacles in the road, such as human beings or animals
- another layer of redundancy in the protection of human life and property. (The
rear taillights can also have this feature.)
15. Proximity
Awareness is something that
every car on the road should have, whether it is a self-driving vehicle, or
not. (Older cars can be retrofitted with a cellular-based device.) Autonomous
cars should always know where other vehicles are in relation to their own
current position, given the variables of speed, distance, and predetermined AOC (‘area of concern’). This
information, along with an awareness of construction, roadblocks, bottlenecks,
and weather conditions, will determine how fast the car travels, and how
closely to other cars immediately around it.
16. A governing
body should be assigned by each municipality (state, city, county, or
country). This body will determine how long vehicles stay in the beta stage of
development, when Instructors and owners (and
their vehicles) must retrain and re-certify, and when cars must be pulled off
the road for maintenance and checkups. Divisions of this governing body would
also exercise authority over other types of autonomous devices, such as drones
and aircraft. (See below.)
17. A designated
and certified owner/operator should be in the vehicles at all times during
the beta stage, and should have the ability to either stop the vehicle, or
seize control of it whenever necessary. Interface with the car can include
voice, control panel, smartphone, and, of course, the traditional
steering/stopping mechanisms.
18. Remote Retrieval
and Operation will allow an
unattended or disabled vehicle to be remotely operated and driven (by PC) into
a service and diagnostic facility, preferably (at first) by a human operator.
19. A.I. Systems and
Diagnostics are essential to
safely operating autonomous or robotic vehicles. The current state and
development of A.I. technology is quickly becoming a mystery, even to its
developers. This
state of affairs must end. If these vehicles are to be re-released safely to
the road, then every decision made by the A.I. system controlling the car must
be understood if it is to be successfully debugged. If a vehicle, or class of
vehicles continually makes the same poor decisions on the road (or, preferably,
on the driving and training range) then the reason(s) it makes these mistakes
must be discovered, and more importantly, understood.
That means knowing what the A.I. is ‘thinking and doing’ all the time. To
assist in this, a series of self-diagnostic routines must be built into the
software. These routines should show variable value dumps and changes, routine
and modular calls, test states, multiple options considered (and why they were
chosen or discarded), and status of the vehicle when (fatal or flawed)
decisions were made. In essence, the A.I. needs its own ‘black box’ built-in, in
conjunction with its programming and decision-making routines, and the
processor(s) for that code. There should also be several levels of diagnostics
(similar to the Star Trek: Next
Generation series.)
20. System Health will be constantly monitored by sensors built
into the vehicle, which will report the status of the cars major systems
(propulsion, brakes, combustion, electronics, environmental, steering, etc.) to
the owner upon starting the car. This information will be gathered during the
diagnostic phase, and relayed to the owner visually (onscreen), and by email,
if the owner desires. It can also be immediately uploaded to the nearest
Control Center or service technician upon request of the owner. (This, for instance, would be requested if
the car were performing oddly or erratically.)
21. Wireless or
Remote Recharging should be the
goal of all vehicles, especially autonomous ones. Such vehicles should be able
to recharge (or refuel) remotely, either wirelessly (while still in motion), at
a charging stations (as some vehicles are doing as of this writing), and by remotely
dispatched flying drones capable of charging several vehicles.
22. Emergency
Navigation: All autonomous
vehicles should be constantly aware of and able to find at a moment’s notice
all emergency service locations, such as hospitals, fueling (or gas) stations,
police stations, and, of course, home.
23. Status Monitoring is a feature of the B.E.A.C.O.N. (see original document) software running on the smartphone or iPhone. Once
removed from its Command Cradle, the handheld device should constantly monitor
the status and location of the vehicle at all times, like an intelligent LoJack
system, alerting the owner if the car is tampered with, vandalized, moved, or
even ticketed and towed. A police or municipality officer ticketing or causing
such a vehicle to be towed would relay this information to his or her precinct,
which, in turn, would report pertinent info to the nearest Control Center,
which will then relay the info wirelessly to the owner.
24. Glass Break
Technology should be
developed (and deployed in a fashion similar to airbags) in the event that a
vehicle is suddenly submerged with passengers trapped inside. The mechanism
should be either manually triggered, or automatically, if the system senses
water has filled the passenger space. Most importantly, it should be configured
and designed in such a way that the shattering glass does not pose an
additional threat to passengers in the vehicle.
25. Fire Control and
Suppression: Each vehicle
should have its own dual part fire suppression system: one for the passenger
compartment (safe to breathe for short periods) and one for non-passenger areas
of the car, such as under the hood, in the trunk, and inside the car’s inner
workings.
26. Authentication
Stream and Authentication Fingerprint is a steady flow
of authentication data, approximately hundreds or even thousands of times per
second, which will help protect the vehicle from unauthorized remote control,
access, seizure, theft, or driving. This stream can be of the highest possible
encryption (or some new type, yet to be developed), and include blockchain
technology. Any significant breach or break in the flow or rhythm of the stream
may indicate that the vehicle’s online security has been compromised. The
authentication fingerprint will be different for each vehicle, based on the maker
of the car, information about the owner, the current trip, the pre-established rhythm of the
data flow, with its own unique pauses and stops, and the home Control
Center alpha-number, several dozen digits and letters long. (Each vehicle will have access to all Control Centers, but call only
one ‘home’ at a time. Home Control Centers can be reassigned, but only by the
governing authority, or by petition to the Authority by the owner.)
27. Co-Pilot: With the dozens of sophisticated systems onboard,
the driver/owner, if he or she so desired, could monitor the steady flow of
information via an onscreen output called the system co-pilot, which would
intercept and manage this data to make it more accessible for human occupants
of the vehicle. It could be configured to share only that information that
would prove to be most vital to the human driver, while withholding the more
mundane (or complex) information unless it
is preconfigured to do otherwise. (The decision and the configuration settings
will be at the discretion of the driver.)
