Machine Learning Limitations Are Getting In The Way Of Autonomous Vehicles

Impressive claims are being made for Autonomous Vehicles (AVs) with Artificial Intelligence (AI) at the controls. There is however a gap between the level of driving ability that these vehicles have achieved to date and that which would pass as fit to drive without supervision. Limitations in Machine Learning systems that are used as the AI software components need to be overcome before this gap can be resolved.

Machine Learning systems contain neural networks of logical pathways joined together by decisions, facts and concepts like the way road networks are joined by interchanges, junctions and roundabouts. They are developed and trained using enormous amounts of data. In the most advanced AVs such as those being tested in U.S. cities the data has been gathered from millions of miles traveled on public roads.

The use of Machine Learning along with data gathered from a variety of on-board sensors has brought AVs up to Level 2 of the scale defined by the Society of Automobile Engineers (SAE) International and shown in the following diagram. At this level, a human driver has control of and responsibility for the vehicle at all times although they can delegate some driving activities. For instance setting the AV to maintain a controlled speed while keeping a safe distance from the vehicle in front.

Although there are vehicles that can claim to have achieved Level 4 or 5 they can only do so in a small set of specific circumstances. Companies in the mass market are struggling to prove that they can safely bridge the gap between Levels 2 and 3 where “the automated driving system monitors the driving environment”. Machine Learning’s strength is that it derives useful information from uncertain data but this is a limitation in situations that require results to be certain and predictable.

In Machine Learning systems the neural network’s logical pathways are created when a training set of data is processed over and over again so that the system configuration converges on the best fit with desired results. Incompleteness of the system’s logical data model has the potential to undermine the accuracy of the system. This affects objectively and accurately assessing the relative success of the results that the system converges towards.

Related to this, the physical data entering the system must be classified, evaluated and prioritised to allow the results from different models to be assessed and compared. The classification scheme and the rating of the outcomes are designed by humans and thus susceptible to both conscious and unconscious cognitive bias.

Two tragic fatal accidents, one involving a Tesla and the other a Uber were caused by failure to take action in response to objects crossing the path of the vehicle, a lorry and a cyclist respectively. The National Transportation Safety Board’s preliminary report on the Uber accident in Tempe, Arizona notes that “the self-driving system software classified the pedestrian as an unknown object, as a vehicle, and then as a bicycle with varying expectations of future travel path.” https://www.ntsb.gov/investigations/AccidentReports/Pages/HWY18MH010-prelim.aspx

The limitations in Machine Learning due to uncertainty introduced by the data can be overcome by systems that provide predictable results. Another post will discuss an approach to the additional systems that will allow the gap to Levels 3, 4 and 5 to be successfully traversed.

Ethics for Autonomous Vehicles

I have confidence that a variety of fully autonomous vehicles (AVs) will arrive on our roads in the next few years. They will have superhuman abilities in vigilance and dexterity but I believe they should behave like machines rather than humans when faced by moral dilemmas.

David Hume with St Giles in the background

David Hume, the great thinker of The Scottish Enlightenment, said that moral decisions cannot be made by reason alone because emotions and instincts are always involved. This is illustrated by taking part in the Moral Machine, MIT’s version of the classic trolley problem thought experiment. It asks for a decision on which of two groups of people should be sacrificed in an accident, a question that rational thought processes are unable to answer.

The implication for AVs is that they are not equipped to evaluate ethical criteria because they are controlled by computer systems based on programming logic. However the thought experiment does highlight potential issues that are worth considering for the design of AVs. These include; “What can be done to guard against the situation occurring?” and “What additional safety features can be designed so that AVs are able to protect people both on the inside and the outside?”

AVs will not be immune to the type of mechanical failure that can cause an accident and if this happens the ability to control the vehicle is likely to be already compromised. When this occurs the best decision might be to take no action and leave the laws of nature to determine the outcome.

The probability of mechanical failure can be greatly reduced by constantly monitoring the performance of all mechanical components and pro-actively arranging for repair or replacement. This will significantly reduce both the potential risk of an accident and the type of disruption caused today by poorly maintained vehicles that have broken down.

However a risk of collision will always exist and the first priority for the design of AVs must be to protect the occupants. It is clear that most people would be reluctant to get into an AV if they thought there was a possibility that they would be sacrificed to save others. This was clearly stated by Christoph von Hugo of Mercedes-Benz last year when he said that protecting drivers and passengers would remain a priority.

Current auto design takes this to an extreme by surrounding the occupants in a hard shell designed to absorb the impact of hitting another vehicle or solid object at speed. If AVs live up to their promise by radically reducing the 75% of accidents caused by human error then there is a case for designing the outer shell to minimise the impact on anything the vehicle runs into.

Although the advent of AVs is expected to reduce the number of deaths and serious injuries, there is one extra area that regulation can address. It should require all available data to be published for serious incidents so that they can be fully investigated by authorised third parties. Any lessons learned must then be addressed consistently by every manufacturer and certified in testing. This opportunity to improve safety by co-ordinated action will reduce accidents even more quickly.

Design for Driverless - Creating A Brighter Now

I’m looking forward to the brighter future that will be ushered in by the advent of driverless vehicles.  It will have far fewer fatalities and make mobility more accessible for the young, the old and the infirm.

All sorts of wonderful scenarios have been suggested for the future that will be delivered by an auto industry that combines mechanical engineering with communications and computing.  However experience has taught me that successful computer systems need to balance ambitious design goals and robust development expertise.  The fantastic potential of a driverless future will only be realised by vehicles that meet the needs of users in ways that are intuitive and easy to use.

To illustrate this let us go back to the questions at the start of the Diversion Use Case discussed in a previous post.  I suspect that most driver’s reaction the first time they are confronted by a diversion sign during a journey is answer No. 3 "I'll carry on and decide whether to follow this diversion when or if I see another sign".  Based on my discussions with other drivers it appears that we treat all temporary road signs with a fair degree of scepticism, relying on a combination of experience and guesswork to choose a pragmatic course of action rather than confidently interpreting and obeying the diversion signs.

Driverless cars will not have this luxury of being selective in the way they choose to behave when they see diversion signs.  They will be programmed to examine each road sign using computer vision and add the instructions it contains to the set of rules that are already in use.  They will also have to do this every time they encounter the diversion and will make decisions based on the information they have gathered up to that point.  If they do not find any They will not know whether the diversion is no longer in operation or if the signs have been moved or knocked over, as happened in the picture at the top.

In the second part of the diversion post a network model with details of the junctions and connecting roads was able to supersede diversion signs as far as driverless vehicles are concerned.  It can also be used to supply data to GPS systems in the present and thus deliver immediate benefits to drivers using GPS systems without having any negative impact.  If such a system is in place when driverless vehicles first start appearing on the road it will help to ensure that they do achieve some of the great potential attributed to them.

I think there will be many other opportunities to start creating the driverless future in the here and now.  Solutions that are designed so that they improve today’s travel as well will be doubly successful by realising shorter-term benefits at the same time as hastening the arrival of driverless vehicles.

The more data the better for driverless cars and cities

The model in the previous blog post is made up of the junctions and the roads joining them, see Fig.1.  This demonstrates that when the junctions are located on a map using appropriate co-ordinates, the presence or lack of a road joining two junctions is sufficient to allow accurate route planning that avoids road closures without having to worry about diversions.

Fig. 1

There is a wealth of additional data that can be recorded about both junctions and roads.  In the case of the roads, the distance travelled between junctions is important.  So is the speed limit, the number of lanes in either direction and whether there are bus or cycle lanes.

In the previous blog post the best available route appeared to be the shortest however this requires negotiating a narrow lane, shown in red in Fig. 2, where vehicles in both directions share the road.

The size of vehicle and the capability of the driver, whether human or automated, might result in a decision to use the alternative route shown in green.  It should be noted that this route is still shorter than following the diversion.

Fig. 2

Additional data can also be recorded about the junctions to differentiate between crossroads, roundabouts and mini-roundabouts.  For junctions that are controlled by traffic lights, an average waiting time can be estimated for each direction of approach to help with route planning.

This data not only means that driverless cars will be able to accurately predict journey times but also provides a baseline for use by the city.  Journeys can be measured for individual cars at different times of day and under different weather conditions to understand when and where congestion occurs.  Potential solutions can then be simulated using the data in the model.

Driverless cars don't need Diversions

What is your reaction to a Diversion sign?

Is it:

1. Thank goodness, I can now be confident of getting round the obstruction quickly
2. I have no idea whether to follow this diversion or not
3. I'll carry on and decide whether to follow this diversion when or if I see another sign
4. I'll rely on the navigation system to handle the diversion and get me to my destination
5. Oh no, this is bound to delay my journey and take me somewhere I don't want to be?

If you answered 1 or 4 then you can entrust your travel to most of the driverless cars currently under development.  However if you answered anything else you will probably agree with me that current diversion signage needs improved if roads populated by driverless cars are to achieve the step change efficiency improvement that has been promised. 

The good news is that it isn't hard to give driverless cars accurate data that will allow them to find the best available route that avoids the road closures caused by a diversion.  Here is an example based on recent roadworks affecting the roads around my home. 

The black route in Fig. 1 is the one I would normally take from the traffic lights at Leith Walk and Annandale Street to my home in Hopetoun Crescent which is a one-way street.

Fig. 1

The road works were for pavement improvements along Leith Walk and removed access to Leith Walk from McDonald Road and Brunswick Road as shown by the star in Fig. 2.  Diversions were put in place on Leith Walk in both directions.  The diversion signs directed me from the Annandale traffic lights to Bellevue Road and then to McDonald Road where the diversion ended.  The diversion therefore took me several streets away from my destination and I had to complete the route shown in blue in Fig. 2 using my own information.

Fig. 2

The Smart Traffic City model represented in Fig. 3 allows in-car systems or central systems such as those used by driverless cars to select the best route without the need for diversion signs.

Fig. 3

The shortest available route for me to use is the one shown in Green in Fig. 4.

 

Fig. 4

Creating a Smart Traffic City improves the quality of data for the growing number of vehicles that that are computerised or connected while allowing current visual information to remain for human drivers.

The model can also hold additional information about roads and junctions that influences route strategy and selection..  This data and the value it provides will be the subject of a further blog posts.

Cheers
 Andy.

What Does Smart Traffic Mean To Me?

City traffic is a complex system that is influenced by many factors like weather, local events and time of year as well as the contents of the streets.  Roads and pavements, signals and signs determine the traffic flow and create an information network that is used to navigate the city.

Driverless cars are learning to drive without a comprehensive understanding of the information that is presented to a human driver.  The delivery of this information can also be improved significantly if it is unconstrained by the ability of human drivers to process the information quickly enough to use it.

Smart traffic in an individual city needs to be open, secure and based on common standards.  Smart traffic systems control signals that optimise individual vehicle journeys to create the most efficient flow throughout.  The objective of a well managed traffic system varies but there are some goals that are common such as; reducing the frequency and severity of accidents, cleaner more efficient propulsion creating a quieter, better quality environment, shorter travel times and less waiting.

Optimising traffic flow while reducing accidents would be a good measurement to assess the success of introducing driverless cars.  Generating and collecting the data to establish a baseline and then track change in both dimensions is an important aspect of any trials.  Every city will be different so an important feature of smart traffic systems is an open architecture that creates standard data without regard to individual deployment choices.

The current priorities of manufacturers of driverless cars suggest that standards for publishing and processing traffic information and instructions may not appear during present trials.  It will be easier to establish standards if the subject gets aired across car manufacturers, transport service operators, city councils and others prior to the start of larger scale roll-outs.

The opportunity offered by combining the ability of driverless cars to process information from sensors, other vehicles and active signage instantly and accurately with smart traffic city systems that can provide information and directions ahead of the decision point must be able to better achieve the dual goals of better traffic flow and a reduction in accidents.

More about the need to know

I am certain that it will be essential for Smart Traffic City Systems to know about intended destination in order to achieve the level of safety and efficiency to which we should aspire. This conviction is partly due to my experience as a motorbike rider.

My personal transport choice is a motorbike currently a Triumph Thruxton (pictured) for reasons of health and pleasure.  There are three things that I do constantly when riding; watch out for cars, assess the road surface and nod or wave to other riders.  Being constantly vigilant and aware of cars in the vicinity is by far the biggest contributor to safety.

Any biker can tell many tales of near or actual accidents caused by a car pulling out in front of them, closing a gap or suddenly changing direction.  I hope a move to driverless cars will not remove motorbikes from the roads and once Smart Traffic City Systems are introduced the advantages of riding a bike will overtake the risks.

A heads-up display like that being developed by NUVIZ using the information from a Smart Traffic City System would display route, digital signage and details of other road users. The key data about other road users is their destination because it can be used to derive routes and thus intended direction at junctions.  I shall still be watching carefully to see what other vehicles are actually going to do but at least there will be a chance that it is what has been predicted.