201
Improving Safety
Pilots with autonomous shuttle buses: From 2017, testing of AVs was allowed
in Norway. In the SmartFeeder ( 2019) research project, initial data are gathered from
five test sites with MAAS pilots. Each pilot tests s elf-driving shuttle buses carrying
up to six passengers, operating at an average speed of 15 km/ h, and with an operator
to monitor and take over control if necessary ( during the test phase). These pilots are
“ fixed route autonomy”, where the autonomous system follows a predefined route and
processes a limited amount of sensor data along the route. The motivation varies,
i.e. solving a last mile problem ( connecting workplaces with public transportation),
testing out technology and user acceptance or property and business development.
In total, the buses in the pilots have driven almost 22,000 km, with approximately
40,500 passengers in both summer and winter conditions. Initial data have been collected regarding disengagement of the system and involvement of the operator in
the pilots in three categories: “ obstacle emergency stop” ( sensors detect something
and automatically stop), “ soft stop” ( operator overtakes system and decelerates the
vehicle) and “ Manual switch” ( for manually driving the vehicle). The collected data
are currently being processed and cleaned for more detailed analysis, and interpretations cannot be drawn yet. However, the reliability and robustness are challenging,
and demands a restricted operating envelope in addition to the need for “ humans in
the loop” when the unanticipated is happening.
Safety Challenges
Tesla with its autopilot has enabled automated driving at high speeds. Several severe
accidents with Tesla autopilot have led Tesla to limit their autopilot functionality. These partially automated vehicle systems at SAE level 2 ( SAE, 2018) always
operate exclusively based on an attentive driver being able to control the vehicle.
For fully automated driving ( SAE level 4 –5), the driver is no longer available as a
backup for the technical limits and failures. Replacing human action and responsibility with automation raises questions of technical, ethical and legal risks, as well
as product safety.
As far as we know from media and public accident reports there have been four
fatal accidents worldwide: three with s emi-automated ( SAE level 2) autopilot and
one with a more fully automated vehicle on public roads ( SAE level 3), the Uber
accident in Arizona where a Volvo refitted with Uber self-driving technology killed
a pedestrian ( NTSB, 2018). In all cases, the autopilot was engaged but without driver
interaction or intervention with vehicle controls, highlighting the need for sensemaking and “ meaningful human control”.
There are few safety records ( data) on SAE level 4 so far. Data from 2009 to
the end of 2015 collected by Google’s cars list three police reportable accidents
in California while driving at 2,208,199 km ( Teoh and Kidd, 2017). This is 1/ 3 of
reportable accidents per km of human-driven passenger vehicles in the same area. In
2017, 19 of 21 reported accidents with G oogle-Waymo cars ( level 4) were r ear-ended
accidents at signalised intersections. This is caused by ordinary drivers’ misinterpretation of automated vehicle behaviour ( as an example expecting that drivers are
not halting when meeting a yellow light at an intersection.). G oogle-Waymo has now
patented a software program allowing their vehicles to drive through yellow light.
A look at accidents and incidents reported to the California Department of Motor
Improving Safety
Pilots with autonomous shuttle buses: From 2017, testing of AVs was allowed
in Norway. In the SmartFeeder ( 2019) research project, initial data are gathered from
five test sites with MAAS pilots. Each pilot tests s elf-driving shuttle buses carrying
up to six passengers, operating at an average speed of 15 km/ h, and with an operator
to monitor and take over control if necessary ( during the test phase). These pilots are
“ fixed route autonomy”, where the autonomous system follows a predefined route and
processes a limited amount of sensor data along the route. The motivation varies,
i.e. solving a last mile problem ( connecting workplaces with public transportation),
testing out technology and user acceptance or property and business development.
In total, the buses in the pilots have driven almost 22,000 km, with approximately
40,500 passengers in both summer and winter conditions. Initial data have been collected regarding disengagement of the system and involvement of the operator in
the pilots in three categories: “ obstacle emergency stop” ( sensors detect something
and automatically stop), “ soft stop” ( operator overtakes system and decelerates the
vehicle) and “ Manual switch” ( for manually driving the vehicle). The collected data
are currently being processed and cleaned for more detailed analysis, and interpretations cannot be drawn yet. However, the reliability and robustness are challenging,
and demands a restricted operating envelope in addition to the need for “ humans in
the loop” when the unanticipated is happening.
Safety Challenges
Tesla with its autopilot has enabled automated driving at high speeds. Several severe
accidents with Tesla autopilot have led Tesla to limit their autopilot functionality. These partially automated vehicle systems at SAE level 2 ( SAE, 2018) always
operate exclusively based on an attentive driver being able to control the vehicle.
For fully automated driving ( SAE level 4 –5), the driver is no longer available as a
backup for the technical limits and failures. Replacing human action and responsibility with automation raises questions of technical, ethical and legal risks, as well
as product safety.
As far as we know from media and public accident reports there have been four
fatal accidents worldwide: three with s emi-automated ( SAE level 2) autopilot and
one with a more fully automated vehicle on public roads ( SAE level 3), the Uber
accident in Arizona where a Volvo refitted with Uber self-driving technology killed
a pedestrian ( NTSB, 2018). In all cases, the autopilot was engaged but without driver
interaction or intervention with vehicle controls, highlighting the need for sensemaking and “ meaningful human control”.
There are few safety records ( data) on SAE level 4 so far. Data from 2009 to
the end of 2015 collected by Google’s cars list three police reportable accidents
in California while driving at 2,208,199 km ( Teoh and Kidd, 2017). This is 1/ 3 of
reportable accidents per km of human-driven passenger vehicles in the same area. In
2017, 19 of 21 reported accidents with G oogle-Waymo cars ( level 4) were r ear-ended
accidents at signalised intersections. This is caused by ordinary drivers’ misinterpretation of automated vehicle behaviour ( as an example expecting that drivers are
not halting when meeting a yellow light at an intersection.). G oogle-Waymo has now
patented a software program allowing their vehicles to drive through yellow light.
A look at accidents and incidents reported to the California Department of Motor