Advancements in Artificial Intelligence for Intelligent Vehicles: A Comprehensive Literature Review

LITERATURE REVIEW

In this literature review , I will discuss previous studies that have been carried out by different

scholars explaining this topic on artificial intelligence in motor vehicle and also an over view on

recent trends and developments in technology.

Intelligent Systems as part of innovation

In this section we will review various objectives which are mainly to analyses the existing

intelligent architectures instead of repeating existing work.

According to Long et al. (2007), a review of intelligent software for autonomous vehicles is

conducted, presenting a set of common used software systems, which we use as a start point for

potential architectures. In his review Long et al. (2007) discusses expert systems (JESS), fuzzy

logic, subsumption architecture, hybrid deliberative and reactive architectures (AURA), through

to the cognitive architectures Soar and ACT. He has not analyzed the strengths and weaknesses

of these approaches discussed, instead they are grouped into a table to be compared in one

location with details such as language, underlying technology, whether they are reactive or

deliberative.

A study by Norling (2012) shows that the suitability of the Belief-Desire-Intention (BDI)

framework is assessed for suitability in modelling human behavior, which highlights the folk-

psychology roots of the framework as a strength in supporting knowledge representation. The

BDI model was put forward which has lead to a variety of software development. It is an

approach applicable for programming intelligent agents, handling the process of sensor feedback,

action selection and processing.

Everts et al (2007) demonstrates the use of a BDI agent in controlling simulated tanks in a virtual

environment, which has relevance to the problem domain of intelligent vehicles considered in

this research. CoJACK is used, a development of the JACK commercial software offering which

has been extended to introduce additional cognitive factors. Although the aim here was to add

more ‘human variability’ to the scenario it highlights that such approaches are feasible and have

proved fruitful in the past.Howden et al (2007) presented a more detailed review of the JACK

which puts JACK forward as an effort to bring the Java flexibility to BDI agent behavior, though

diverging from the more AI specific programming approaches. The proposed “Simple Team”

model introduces the capability of team-based reasoning, although this appears to be more of a

mechanism to specify team behaviors and organization than a distributed mechanism to improve

shared situational awareness.

According to Shendakar et al (2008) the use of BDI agents is presented in an emergency

response and crowd simulation scenario. BDI agents are used to represent individuals,

specifically how they react to danger threats and choose an exit path.

Their conclusion following this discussion was that extended BDI model could be used in

driverless cars is also relevant to the problem domain we are interested in, although no specific

implementation plans are proposed in their work.

Artificial Intelligent Vehicles

In discussing the artificial intelligent vehicles, we will first consider recent developments in the

area of intelligent vehicles to be informed of what such capabilities may look like, what they may

enable, and how we can adopt an approach good to interface such vehicles.

Early studies in this field was driven by the DARPA. These events aimed at pushing the art of

the possible in order to advance capability in this area. The 2005 competition was won by the

‘Stanley’ entry from Stanford, based on a modified Volkswagen Touareg. There are novel

approaches adopted in their approach, discussed in Davis (2006), where developing awareness

and reasoning about the vehicles own perceptions has proved the most fruitful, and this improved

reasoning is put forward as being able to offer improved handling of sensor noise.

Focusing on the AI implementation of their solution as discussed in the “Behavior Generation”

engine has been implemented as a state machine, using high level behaviors which then have

supporting sub-behaviors. Their “Motion Planning” aims to avoid collisions with static and

dynamic objects, and seems to be concerned primarily with road following. The “Perception and

World Modelling” seems to compromise a data fusion engine, driven by a map of both static and

dynamic obstacles. We reflect that an approach of breaking required functionality into

components, with an appropriate level of knowledge representation, proved an effective solution.

Within Tartan Racing’s “Perception and World Modelling” component substantial data fusion

takes place, with base level sensor feeds being fused and classified to provide some “Situation

Assessment” (their term). In the fusion process an object is classified into a more explicit type, if

possible, (e.g. change a set of detected points into a ‘car type’ if it meets the requirements of

being a ‘car’). Regarding “Situation Assessment”, it is stated that the “layer attempts to estimate

the ‘intention’ of the tracked object by integrating the estimates with knowledge about the road

world model”. It is also reported that the system struggles to perform well when approaching

intersections and projecting future events (e.g. whether a vehicle will leave or join at that

intersection). This is another example of how communication of higher level information

between vehicles could prove useful; rather than having to rely on some visual cue (e.g. an

indicator light), vehicles would have been informed as to what was likely to happen at that

intersection based on other vehicles exchange of future plans. As well as having the benefit of

aiding autonomous vehicles in working together, there is the additional benefit that excessive

braking and acceleration are reduced as vehicles are able to predict events rather than relying on

last minute reactions.

Additional discussion relating to the ‘Tartan Racing’ entry is provided in showing details of

potential benefits of automated vehicles. Worth noting is the limitations placed on the scenario

that this vehicle was used in; no pedestrians were present, traffic lights were limited to red only

and at pre-defined points. We draw on this as a reminder, that if an overly complex scenario is

used then our framework may fail; we need to consider a balance between over simplifying the

problem domain versus adding too much complexity.

Two individuals involved in ‘Grand Challenge’ entries, Sebastian Thrun and Chris Urmson,

went on to become involved in the Google Car (e.g. discussed in article resulting in a fleet of

Toyotas which have now covered (as at 2012) more than 190,000 miles (figure from and a recent

blog post by Urmson stating that as of 2014 the have now logged 700,000 miles. Setting out to

achieve the goals of “Reducing road accidents, congestion, and fuel consumption” provides us

with some indication of potential challenges we could seek to address in our test scenarios.

Other manufacturers are making similar developments, with recent announcements such as

Nissan’s Gordon- Bloomfield (2006) further demonstrating the spread of such capability. We

draw on this to increase confidence that assumptions regarding technological capabilities made in

this thesis are close to (if not already) in existence. Physical sensors are capable of providing

increasingly accurate geospatial and environment information, to an intelligent control system,

capable of generating appropriate commands in order for the vehicle to achieve its goals. With

such individual vehicle capability emerging, a question follows as to what communication

between vehicles would be possible and what information could exchange for individual or

group benefit.

Considering alternative autonomous transportation, the development of Urban Light Transit

(Ultra) presented in Low son (2003) as a means of vehicle transport has been successfully in

place in London’s Heathrow Airport since 2011. In this approach, light weight vehicles have

their own guide way constructed with vehicles continuously moving around the route. The

overall network is considered as a number of ‘slots’ with each vehicle then filling a ‘slot’ with an

appropriate gap ahead based on speed and braking. Vehicles are able to merge back into these

‘slots’ on picking up passengers and continue along the route. An empty vehicle management

system is used to address the availability of vehicles in order to reduce wait times, planning

vehicle availability around demand by managing where to park vehicles when not in use, and

minimizing delivery time to pick up points. This demonstration of a level of coordination

between autonomous vehicles is taken as a positive indication of what is possible, and although

the traffic network operates segregated from human drivers, it is still dealing with the

challenging problem of controlling real world vehicles, handling their sensor data, and planning

future maneuver’s.

Artificial Intelligent Vehicle Communication

Communication between vehicles is named as Vehicle to Vehicle (V2V) and

Vehicular ad hoc networks (VANETs), considered as a subset of the Mobile adhoc network

(MANET) family. Developments in this area have been supported by availability of required

technology (e.g. emerging IEEE 802.11 standards and routing protocols such as those reviewed

but also as vehicles become able to measure and capture useful information (e.g. sudden

prolonged use of brakes) which would be of benefit for other vehicles to be made aware of. The

US National Highway Traffic Safety Administration (NHTSA) announcement Nailor (2014) that

Vehicle to Vehicle (V2V) communication devices may become mandatory in a year adds some

weight to the assumption that such communication will be seen increasingly in the future. Other

development is that of ‘Floating Car Data’, where information such as vehicle speed, location,

and direction is reported back to some other system, in order to assist with activities such as

congestion monitoring. Such information can be extracted from the vehicle itself as detailed in

Huber et al (1999), or alternatively via a mobile phone onboard the vehicle. We have an interest

in such developments, as it indicates both the near availability of rich data and also some domain

specific benefits (collision avoidance, congestion reduction) where such information can be of

benefit.

This work is part of the EC-funded “Safe Road Trains for the Environment” (SARTRE) project,

whose aim is to develop the capability to allow a number of autonomously controlled vehicles to

follow one human-driven vehicle. Publications from this project identify the required

functionality to control convoys, as well as some consideration for what may be communicated

(control and coordination of the platoon). This work also provides examples of the potential

benefits of vehicle platoons: up to twenty percent reduction in fuel consumption, ten percent

reduction in fatalities, and improved driver convenience (for passenger-drivers in the vehicles

where control has been ceded to the platoon). Further benefits have also been discussed in where

improving traffic efficiency is a key goal.

Widening the scope from vehicle to vehicle communication, there is ongoing research in vehicle

to infrastructure (V2X, V2I), which also has the potential to enhance individual SA (I.e. a vehicle

is informed of an upcoming traffic light state).Kim (2011) has adopted such an approach, and

demonstrates benefits in implementing communication between traffic lights and vehicles, in

order to improve fuel consumption and reduce emissions. Similarly, Audi (published via parent

company website have demonstrated vehicles retrofitted with a device to allow them to interact

with traffic lights, that appears to show real benefits for traffic flow. Specific improvements cited

include a reduction of CO2 emissions by up to 15 percent, and substantial (sic) fuel savings, but

the lack of precise details makes verification difficult.

This study shows that if navigation systems share traffic information, then journey times are

shortened, and we draw on this, as well as the previous examples to conclude there are

significant potential benefits from improved communication in the vehicle domain. We suggest

this may be due to both improving the collective SA, but also as it enables better coordination.

Lastly to inform our work in the area of how much communication may be possible based on

such technology, the capability of such networks needs to be assessed, in terms of metrics such

as transfer speeds, latencies, packet loss etc. There is an inherent difficultly in the physical nature

of establishing a moving networks. As such, it seems reasonable to expect performance to be

lower than that of a stationary 802.11b/g/n wireless network. In Jiang et al (2009) a transmission

speed of 6 Mbps is considered, however alternative communication strategies such as cellular

approaches (e.g. et al (2012) may offer improvements on this. In short, this is a rapidly

developing area, and what communication performance is available now may be considerably

different in a few years’ time.