15
he concept of a car that can drive itself is actually not a
new one in the scientific community: as early as the 1980s,
there were autonomous cars able to operate in very simple
environments at relatively low speeds. In recent years, Google’s
driverless car project has tested a fleet of cars in Nevada, Florida and
California with the team completing 300,000 miles of autonomous
driving on public roads as of 2012.
The self-driving car promises a number of benefits with the potential
to improve the safety, efficiency and accessibility of our roads. Safety
will be improved because the on-board computer will not get tired
or become distracted by a ringing phone. An autonomous car would
also operate more efficiently, reducing fuel consumption, pollution and
saving money. By allowing autonomous vehicles to communicate it is
possible to create convoys with greatly reduced gaps between each
car, effectively increasing the capacity of the road network without
the need to widen existing roads or build new ones. Furthermore,
an autonomous car could free up those wasted hours stuck in a car
on the commute to work. Potentially the most exciting benefits are
those provided to the elderly and disabled, for whom technology could
provide some of the independence and freedoms that they have either
never enjoyed or thought they had lost.
With all of these potential benefits to be realised, in October
2012 the Oxford Mobile Robotics Group took delivery of
two all electric Nissan Leafs. Over the next five months one
would become capable of driving itself. The research group
led by Professor Paul Newman, BP chair at Keble College, and
Dr Ingmar Posner started to develop the full set of systems
required to enable a car to operate autonomously. Working
with colleagues from Nissan, who were embedded in the
research group, the vehicle was modified such that the onboard
computers could control the car. Three software systems
were developed to enable autonomous operation: navigation,
determining the position and motion of the vehicle; perception,
generating an understanding of world around the vehicle; and
planning, calculating the path the vehicle should drive.
Up to now, the majority of autonomous systems have been fitted with
a very expensive high performance inertial navigation system (INS),
which uses a combination of GPS, accelerometers and gyroscopes. The
benefit of an INS is that it can measure the position of the vehicle with
an accuracy of a few tens of centimetres. However, these systems
only achieve this level of accuracy in conditions where the GPS is
working well. When driving under a tree canopy or in a built up area,
the view of the sky is obscured and the performance of these systems
can degrade. For this reason the major challenge to the team was
to develop a system which was not reliant on GPS and instead uses
comparatively low cost sensors.
The solution developed by the group was designed to use the existing
structure of the surroundings, therefore only requiring that a vehicle
had driven the road previously and built a map of the environment.
The navigation system then makes observations of the local
surroundings and matches these to the prior map to determine the
position of the vehicle. The task of mapping all roads appears to be a
daunting one. However, Google’s Street View project has shown that it
is not unrealistic.
The prior information used by an autonomous vehicle need not only
be maps or images. The focus of my work will be to look at how the
planning system can utilise observations of the way in which other
vehicles have driven. By learning from experience, the best way in
which to drive each road can be determined, resulting in a vehicle
which will approach a junction such that the view is not obscured
where there are often parked cars. This means that your self-driving
car will become more capable the more it is used. The vision for this
work is to allow these experiences to be shared, which will enable the
vehicle to operate in places you have never been before with all the
experience of a local driver.
Although a fully autonomous vehicle is unlikely to become available
to consumers until at least 2020, if one were to look at the features
available on modern cars it becomes evident that your car is actually
already closer to full autonomy than you might think. Anti-lock braking
systems and traction control intervene in the event of a skid, while
parking sensors and auto-parking allow the driver to parallel park
controlling only the accelerator and brake. Adaptive cruise control
systems not only maintain a constant speed but also adjust the speed
to match that of the car in front, with modern systems able to detect
road markings and steer the vehicle to follow the lane. Auto-brake
systems can perform emergency stops to avoid collisions with
pedestrians and other cars, and by using a range of sensors such as
laser, camera and radar these systems operate even at night. Whilst it
is unlikely that your next car will be one you never drive yourself the
progression towards the goal of autonomous cars is well on its way.
Geoff joined the Oxford Mobile Robotics Group in 2012 as a DPhil
student. Prior to joining the MRG, Geoff was a Senior Scientist in
the Information and Intelligent Solutions theme at the BAE Systems
Advanced Technology Centre focussing on motion planning for
autonomous vehicles.
Geoff Hester
2012 Engineering Science DPhil
t
The Robot Car
Not many years ago the idea that your car would be driving you to work was confined to the realms
of science fiction. With advances in the processing power of modern computers and the application
of information engineering we are significantly closer to realising the self-driving car.
Autonomous car uses a laser scanner to build a 3D map of the surroundings
