Call for Papers - African Journal of Biological Sciences, E-ISSN: 2663-2187, ...
Autonomous Vehicle Benefits and Challenges;Role of Governments?
1. Autonomous Vehicle
Benefits and Challenges;
Role of Governments?
Christopher A. Hart
Presented to
Ford School Center for Local, State, and Urban Policy
February 10, 2020
Hart Solutions LLC
3. Opportunities
⎼ Automation has generated significant benefits in
aviation
• Safety
• Aircraft productivity and operating efficiency
• System throughput
• Reduced pollution
⎼ Benefits from driverless cars will probably include all of
those – especially safety – and many others
⎼ Issue is how best to avoid exacerbating public
skepticism about automation by minimizing injuries
and damage in the development process
February 10, 2020 Hart Solutions LLC 3
4. Aviation Automation Lessons Learned
⎼Importance of “Human-centric” automation
⎼Automation not perfect; thus, important to
have “Graceful exits” if:
• Driver is inattentive
• Automation fails
• Automation encounters unanticipated
circumstances or is uncertain
⎼Humans are not good monitors of reliable
systems
February 10, 2020 Hart Solutions LLC 4
5. “Human-Centric” Automation
⎼Aviation automation began with “Automate
because we can”
⎼Outcome not always optimal; evolved to “human-
centric” automation
⎼Many AV efforts to be “human-centric,” but crashes
reveal need for improvement
February 10, 2020 Hart Solutions LLC 5
6. Not Adequately Human-Centric
⎼Calling system “Autopilot”
• Misleadingly suggests no need for driver engagement
⎼Expecting driver to remain attentive, even with very
effective systems
• Effective systems encourage complacency, inattention
• Responses to inattention – stopping car on the road or
disengaging automation
• Both responses have serious flaws
⎼Without “graceful exit” re driver inattention, ready for
“prime time” on the streets?
February 10, 2020 Hart Solutions LLC 6
7. Other Human-Centric Realities
⎼ If not adequately user-friendly, automation may
• Be improperly used
• Distract driver
• Cause driver to turn it off, lose potential benefits
‒ Possibility of automation being uncertain
• Inadequate lane markings
• Less braking effectiveness, e.g., slippery road
• Unanticipated circumstances, e.g., object in the road
‒ Adequacy of alert to driver if automation fails or is
uncertain?
February 10, 2020 Hart Solutions LLC 7
8. Automation-Generated
Complacency – Williston, FL, 2016
− Car submarined under trailer of
left-turning tractor-trailer
⎼ No car skid marks
− Driver inattentive, using “Autopilot”
− Owner’s Manual: Use
Autopilot “only on highways
and limited access roads”
− Driver over-reliance on
Autopilot; Tesla over-reliance
on Owner’s Manual
February 10, 2020 Hart Solutions LLC 8
9. Complacency (con’t) –
Culver City, CA, 2018
February 10, 2020 Hart Solutions LLC 9
⎼ Fire truck stopped in
left lane of Interstate,
responding to crash
⎼ Tesla also in left lane, car
in front of Tesla moved right to avoid fire truck
⎼ Tesla, apparently in automated mode, remained
in left lane and struck the truck
⎼ NTSB investigation underway
10. Automation Failure, No Alert
⎻Fort Totten Metro, Washington, DC, 2009
• Train became electronically
invisible, no warning to
following train
• Following train sensed no train ahead, began
accelerating to max speed for that area
• Operator applied emergency brake upon seeing
stopped train, but her sight distance was limited by
curve
• Operator, eight passengers killed
February 10, 2020 Hart Solutions LLC 10
11. Unanticipated Circumstances
⎼ Air France Flight 447, Rio to Paris, 2009
• Cruise at 35,000’, night, near
thunderstorms, autopilot engaged
• Ice blocked pitot tubes, thus no airspeed
information
• Autopilot and several other systems
inoperative without airspeed information
• Pilot responses inappropriate, crashed into ocean
‒ Query: Pilot training re
• loss of airspeed information in cruise
• manual flight at cruise altitude
• recovery from stall/unusual attitudes at cruise altitude, or
• crew resource management?
February 10, 2020 Hart Solutions LLC 11
12. Automation Challenges
Not Encountered in Aviation
⎼AI learns with experience
⎼Drivers not trained
⎼Street testing is essential
⎼Need for “Graceful exits”
⎼Mixing driverless vehicles with human drivers
February 10, 2020 Hart Solutions LLC 12
13. Automation Challenges
Not Encountered in Aviation (con’t)
⎼Software updates: Ensure no undesirable interactions
between systems or unintended consequences?
⎼Cyber security concerns, initial and with invasion
protocol improvements?
⎼Competition re safety
⎼Ethical issues
February 10, 2020 Hart Solutions LLC 13
14. Automation That Learns
⎼Aviation automation does not learn; changes
necessitate additional pilot training
⎼AV automation will learn with experience
⎼Automation changes from learning create challenge
of keeping drivers abreast of automation behavior
February 10, 2020 Hart Solutions LLC 14
15. Driver Training
⎼Airline pilots have extensive training, including when
systems are modified
⎼Designers of automation that assists drivers must
assume worst case, i.e., drivers will have no training
and will not look at owner’s manual
⎼Possible outcomes from untrained drivers:
• Driver turning automation off, losing protection
• Driver being distracted by automation
⎼Will auto dealers need to train buyers?
February 10, 2020 Hart Solutions LLC 15
16. The Necessity of Street Testing
February 10, 2020 Hart Solutions LLC 16
⎼ Street testing is essential
because the streets are so
complex and variable
⎼ Lab and test track
testing are not sufficient
17. The Inherent Challenge
⎼Assuming vehicles are very reliable before being
tested on the streets, the challenge is that humans
are not good monitors of reliable systems
⎼Needed:
• Improved system design, e.g., warning re uncertainty
• Improved monitor training
February 10, 2020 Hart Solutions LLC 17
18. Human as Inadequate Monitor:
Tempe, AZ, 2018
February 10, 2020 Hart Solutions LLC 18
‒ “Driverless” street test (with
monitor/driver)
‒ Woman killed walking across
street; night, not in crosswalk
‒ NTSB investigating, preliminary
report notes no pre-impact slowing
‒ First pedestrian fatality from “driverless” vehicle
19. Need for “Graceful Exits”
⎼As with systems that assist drivers, need
“Graceful exits” without drivers if:
• Automation fails
• Automation encounters unanticipated
circumstances or is uncertain
⎼Graceful exits:
• Important when driver is being assisted
• Essential when driver is removed
February 10, 2020 Hart Solutions LLC 19
20. “Graceful Exit” re
Landing on the Hudson, 2009?
−Both engines ingested birds
−Pilots unable to reach airport,
landed in river
−No fatalities or serious injuries
−Could automation have landed successfully?
February 10, 2020 Hart Solutions LLC 20
21. Mixing Driverless With Humans
⎼Very challenging because of variability and lack of
predictability of human drivers, bicyclists, and
pedestrians
⎼Necessitates consideration of human factors of other
humans, even for full (driverless) automation
February 10, 2020 Hart Solutions LLC 21
22. Software Updates
⎼As updates are broadcast – testing re:
• Interactions or unintended consequences?
• Driver responses to updates?
⎼Aviation system designers use pilots during design to
test the human-machine system
⎼Using drivers in auto simulators is less effective due to
• Less sophisticated simulators
• Greater variability of background and experience of drivers
February 10, 2020 Hart Solutions LLC 22
23. Cyber Protection
⎼Robustness of testing update broadcast process and
updated systems against cyber attack
• Initially
• As attack protocols continually improve
⎼Robustness of testing cyber protection updates
• Re unintended consequences
• Against continually improving cyber attack protocols
February 10, 2020 Hart Solutions LLC 23
24. Competition re Safety
⎼Airlines do not compete on safety
⎼Lack of competition enables free sharing of info re
safety issues and remedies
⎼Automakers compete vigorously on safety
• Eliminating safety competition is unlikely, probably
undesirable
• Challenge – exploiting competition while enabling sharing
of safety information
• No competition re objectives, but compete re most
effective implementation?
• Compete to be safest rather than first?
February 10, 2020 Hart Solutions LLC 24
25. Ethical Issues
⎼ No major automation ethical issues in aviation
⎼ Overarching ethical issue
⎼ Introduce AVs sooner (systems less perfected, but begin saving
lives sooner) or wait until systems more perfected (thus suffering
more fatalities)?
⎼ Other ethical issues
• Save occupants or save others?
• Select response that minimizes fatalities?
⎼ Better to address ethical issues early rather than as after-
the-fact add-ons
⎼ Ethical issue activity mostly from academia
⎼ Federal leadership needed
February 10, 2020 Hart Solutions LLC 25
26. Role of Federal Government
⎼National uniformity
• Avoid patchwork quilt of requirements from
state to state
⎼International harmony
• Simplifies international trade
⎼Ethical Issues
February 10, 2020 Hart Solutions LLC 26
27. Sample State and Local
Government Issues
⎼ Driving Requirements (if Feds don’t establish
requirements)
• Licensing requirements
• Hands on the wheel
• Requirements re steering wheel, brakes
⎼ Infrastructure
• Traffic signals
• Lane and center lines, parking lines, crosswalks
• Street signage
• Street parking
• Parking garages
• Dedicated lanes for AVs
• Segregation of bicycles, motorcycles, pedestrians
• Communication with and between AVs
February 10, 2020 Hart Solutions LLC 27
28. Sample State and Local
Government Issues (con’t)
⎼ Revenue
• Auto registration
• Fuel taxes
• Parking and moving violations
• Parking lot taxes
⎼ Resources
• Traffic police
• Emergency response
• Infrastructure maintenance
⎼ Other
• Differences for bad weather?
• Reduced organ donations?
February 10, 2020 Hart Solutions LLC 28
29. Conclusions
⎻ Automation offers significant potential benefits,
including saving tens of thousands of lives every year
in the US, more than a million every year worldwide
⎻ However, there are many challenges
⎻ Learning lessons from decades of aviation automation
experience can help avoid exacerbating public
skepticism re automation, avoid unnecessarily
delaying implementation
⎻ AV industry will also face many additional automation
challenges
⎻ Anticipate major changes for governments at all levels
February 10, 2020 Hart Solutions LLC 29
74 mph, no skid marks
Car didn’t see truck, can’t spot crossing objects
Driver not paying attention
Name ”Autopilot” misleading
Bragging on social media, Sudoku
Driver engagement: steering wheel torque?
Many warnings, all ignored
Owner’s Manual: Complied? Who reads?
Train electronic disappearance
Difficult to replicate
Parasitic oscillation
Warning on dispatch board: 500/day, ignore overalarm
No warning to following train
Several sudden warnings
Inconsistent – too fast, too slow, no cause and effect
Startle
Desired response, with hindsight – pitch and power
Actual response – immediate full nose up by copilot . . . Not sure why
No CRM, pilot didn’t know about copilot’s nose up
Advisability of scheduled break?
Three pitot systems – redundant if taken out by common cause?
Yoke vs. sidestick
2000/1000 --- necessity for autopilot
No sim training re manual flight at cruise
No sim training re loss of airspeed in cruise
No sim hi altitude stall recovery training
False stall recovery alarms; based on airspeed, not weight on wheels
Previous loss of airspeed info events – pilots recovered, so better heater, but no emergency
Need to consider automation plus human and automation plus system
Streets and highways are so variable and complex that street testing is essential; lab and test track testing are not sufficient
Sims not sufficiently robuts, more driver variablilty
Pilots can get certified from sims alone, no real airplane time
Assuming vehicles are very reliable before street testing, challenge is that humans are not good monitors of reliable systems
Needed: Improved system design and monitor training
Full disclosure – engaged by Uber
Crash inevitable
Need street testing, but humans not good monitors of reliable systems
Industry issue, not just Uber issue
With such extensive street testing, inevitable to have distraction, not necessarily for bad reason, plus victim sooner or later
Inadequate system warning re uncertainty
Stop in time if looking? NTSB will determine
Pedestrian impaired
No crosswalk where sidewalk meets street
Bird ingestion anticipated
Good news – ingestion contained
Ingestion in both engines not anticipated
Pilots not trained to glide, land in water, or land without power
Could design recovery automation in hindsight, not in foresight
Automation issue – phugoid damper denied 3 ½ degrees nose up in flare; more damage?