When AV Software Meets Human Driver…

A key question about autonomous vehicle (AV) technology is whether an AV software driver can match the capabilities of human drivers — and if so, when will this happen?

We know that developing a safe AV software driver is extremely challenging. There is general agreement that the best approach is to deploy AVs for relatively simple use-cases first — slow-speed goods AVs, hub-to-hub autonomous trucks, fixed route and low-speed AVs and local area robotaxis.

The vast majority of human drivers are quite good — especially when they are focused on the driving tasks and have multiple years of driving experience. However, most drivers are not fully concentrating on the driving tasks part of the time. Or young drivers have not yet reached the skill levels to hand difficult driving events. And older drivers eventually lose some of their driving proficiency.

The result is an unacceptable level of car crashes across the world. Total road deaths per year has topped 1.35M according to United Nation statistics. U.S. road deaths have ranged between 35,000 and 40,000 in the last five years.

Another sobering data point is the total cost of all vehicle crashes in the U.S. The last detailed study by U.S. Department of Transportation was done for 2010 and released in a 2015 report. This report showed total crash costs in 2010 reached $242 billion, or 1.6% of GDP, or $784 per person. When quality-of-life valuations or total societal harm costs are included, total crash cost was $836 billion or equivalent to 5.5% of GDP. Societal harm costs include quality of life valuations from loss of life and severe injuries. The details of the cost calculations are documented in the linked report.

A major reason for developing autonomous vehicles is to lower the cost and impact of car crashes. Human driving errors are by far the leading cause of car crashes and accounts for over 90% of total crashes. The big question is whether the best AV software driver can lower the crash rate compared to human drivers — and if this happens — how long will it take?

This column will give some perspectives on how AV software drivers stack up against human drivers. The next table lists key driving issues and compares how AV software drivers measure up versus human drivers. Most of the issues relate to known problems that most drivers struggle with part of the time. A few AV software driver issues are also included.

It is important to remember that there is a large difference in driving skills among human drivers. There is also a large difference in how far the AV software driver development has come. I will use Waymo data when numbers are needed or estimating future trends.

Click the table to enlarge. (Source: Egil Juliussen)

Driver’s license
The driver’s license system, with a written test for traffic rule knowledge and a driving test to show car operation skills, is standard for human drivers before they are allowed to operate a car. There are variations on how strict the tests are between U.S. states and other countries, but the system works.

The traffic rule knowledge is embedded in the AV software driver and this software will be updated on a regular basis. There is currently no equivalent procedure to the driving test for the AV software driver. This is an issue that needs an answer and it is unclear when and how a solution will be available. Currently AV testing is controlled via a permitting system with various degrees of competency checks.

Driving experience
Driving experience is the key to becoming a competent and safe driver, for humans and AV software. For human drivers the driving skills versus years of driving looks like a wide bell curve. The human with a new driver’s license is at the beginning of the driving skill bell curve and moves up the curve with more driving experience. As the human driver ages into senior phase of life, the driving skill declines and eventually the driving skill are not good enough to retain a driver’s license.

The driving skills curve for AV software driver is similar to penetration growth curve, but we don’t know what key parameters are. We know that driving skills for AV software has grown tremendously, but still has a long way to go. The climb up the curve relates to both road miles and virtual (simulation) miles but depends on a variety of AV technologies such as sensors, AI, software and processing capabilities.

Distraction
There are three distracted driving factors: visual (eyes off the road), manual (hands off the wheel) and cognitive (taking your mind off driving). Distracted driving is a major cause of crashes by human drivers and accounts for about 18% of all U.S. crashes and 10% of fatality crashes.

The AV software will not have any distractions. That said, it remains possible that sensors will have issues in classifying objects, and that might be considered similar to human drivers’ visual distractions. As the technology of sensors advance, this will cease to be a factor for the AV software driver, however.

Speeding
Speeding is a leading cause of human crashes and accounts for 32% of fatality crashes and about 20% of all crashes. Speeding will not be allowed in the AV software driver and should not be a factor in AV crashes.

DUI
Driving under influence includes two categories: alcohol and drug impaired driving. Crashes from alcohol impaired driving have been tracked for 40+ years. Drug impaired driving has little data available, but NHTSA has started investigating this problem.

Alcohol impaired driving remains a major factor in causing crashes, but it has declined over the last 35 years. In the last five years alcohol intoxication has been a factor in around 20% of all crashes compared to around 35% in 2010 and over 50% in mid 1980s.

Impaired driving is not applicable in AV software driver. The closest to impaired driving is successful cyberattacks. Cybersecurity protection will be a difficult issue for AVs and some success is likely. If the auto industry implements the emerging  cybersecurity standards, this should not be a factor.

Reaction time
The reaction time of human drivers depend on driving experience and many individual factors. Driver distraction issues will lower the reaction time of human drivers. The AV software driver will have better reaction time than human drivers due to more sensors that have a 360-degree view and computing speed.

Tiredness
Getting tired while driving is a common problem for most drivers — at least part of the time. This will never be a problem for AV software driver. There is little data on crashes where the cause is being tired.

Weather
Most human drivers can handle a variety of weather situation and is currently better than AV software drivers. The main problem for human drivers is being over-confident in bad weather. Over-confidence often shows up on judgement whether to drive or not and not slowing down enough. This happens in snowy weather and flooded roads and too much speed in fog.

AV software drivers are current fair-weather drivers and need much more training to match the average human driver in bad weather. But AV software drivers should have better judgement on whether to drive in inclement weather.

Edge cases
To minimize edge cases is the key to advance the driving skill of the AV software driver. An edge case is when the AV software driver encounters a new driving situation and does not know how to navigate this situation. Ideally, all the edge cases have been learned or they are extremely rare, but AV software drivers are not there yet. When we get there will depend on AV use-cases.

Edge cases is an advantage for human drivers. Human drivers can handle edge cases by using common sense and extending existing driving knowledge. Human drivers also know how to communicate with other road users — usually via simple hand signals.

Crash avoidance
Crash avoidance depends mostly on the driving skill and experience of the human driver. The story is the same for the AV software driver— skill and experience. The human driver must also avoid three items that dull driving skills — distractions, speeding and impairment.

The AV software driver need three items that will minimize system failure. The AV software driver needs a fail-soft architecture, which also called graceful degradation. The hardware that the AV software runs on must have hardware redundancy at every level. When the AV software driver meets another edge case, teleoperation by a remote human driver can be a backup.

Future questions
There are a few interesting questions on future trends. A lot of new cars have ADAS functionality that helps the human drivers be alert via warnings and/or automate some simple driving functions. Will this trend lower the crash rates of these ADAS vehicles? Early data indicates that lower crash rates are emerging for some of these ADAS vehicles. The next question is whether these ADAS skills might dull future driving skills as human drivers do not get to practice some of their skills.

The number of senior or older drivers are increasing in most countries. There is clear evidence that most drivers older than 70 years slowly lose part of their driving skills. These drivers also drive less miles with age, but will this trend have a negative impact on car crashes?

There are also future questions for AV software drivers. How long safety drivers are needed depends on the AV use-case and if teleoperation can be a backup. AV software driver will need to communicate with other road users. Readable displays on the outside of an AV is technology that is emerging. V2X is another useful technology — especially V2P (vehicle-to-pedestrian) since most smartphones are likely to have C-V2X capabilities long before most cars.

The most important AV question is how quickly will edge cases be learned. We don’t have clear answers and that is why easy AV use-cases are emerging first since they have fewer edge cases.

Summary
The crash statistics of human drivers are quite revealing and shows that three driving issues account for about 58% of all crashes in the U.S. The three human driver weaknesses are distractions, speeding and DUI. The human drivers are quite good at dealing with edge cases.

It is the reverse for AV software drivers. The three human driver issues have no impact on the future crash statistics of AV software drivers. Yes, future AV will have crashes, but the goal is that the crash rates are eventually much better than human drivers.

The most likely crashes for AV software drivers will come from edge cases and hence the focus will be to learn as many edge cases as quickly as possible.

Key question: How much social cost reduction must Software Drivers provide versus Human Drivers’ current social and crash cost? This includes ADAS improvements for human drivers.

The post When AV Software Meets Human Driver… appeared first on EE Times Asia.

from EE Times Asia https://ift.tt/30CvKBz