By Colin Barnden
What’s at stake:
Robo-drivers don’t understand how to interact with human-driven vehicles and road safety isn’t improved by autonomous developers blaming humans for crashes involving autonomous vehicles.
Why is autonomous driving on public roads so hard to crack? Why is it taking so long? And why has the U.S. National Highway Traffic Safety Administration (NHTSA) opened an investigation into GM Cruise’s robotaxis operating in San Francisco?
Let’s step back from the autonomous hype, hoopla, and hysteria of the CES circus and take a back-to-basics look at four key areas that provide insight into why robotaxi and robotruck developments have hit the skids.
We can use two familiar examples to easily demonstrate the real-world challenge of prediction. For coin tossing, the odds of a coin landing as heads or tails are exactly equal, which is why this method is frequently used to settle disputes and bets.
How about dice rolling? If we take six, six-sided dice and roll them, what will be the outcome? We don’t know because the outcome is uncertain. We know there will be an outcome and we can estimate the probability of every outcome, but we cannot precisely predict the outcome. This is the effect of uncertainty.
Now answer the following questions:
What is the sensor suite and quantity of compute processing necessary to predict coin tossing? How does it compare to dice rolling? If a simulator is designed to perform coin tossing or dice rolling one trillion times per second, how long does it take to build the necessary dataset and train a machine to predict outcomes without uncertainty?
Recommended: CES Robocar Panel: ‘Humans Are in the Loop’
As humans we know that every toss and roll is independent of every other toss and roll. The past doesn’t predict the future, no matter how many times the process is repeated. If you doubt that, take all your money to the Las Vegas craps tables and see how you fare.
Coin tossing and dice rolling are simple examples for exposing a fundamental challenge facing autonomous development: Perfect knowledge of every factor in a system does not perfectly predict outcomes of that system.
A human baby is born capable of recognizing her mother’s voice. This tells us that humans begin to develop perception in the womb. Sight, sound, smell, taste, and touch are the best-known senses. Proprioception, sometimes described as the sixth sense, is the sense of self-movement, force, and body position, while thermoreception is the perception of temperature.
Our senses and perception serve many vital purposes, including the classification of threats and risks. Relative speed, mass, density, and proximity all alert us to physical risks. Smell can alert us to the threat of noxious gases or fire; taste of rotten or undercooked food.
Humans perceive situations using sense and feeling, as well as knowledge-based reasoning. A six-foot chicken walking beside a three-foot chicken is probably a parent and child heading to a costume party. That “80” miles per hour speed sign on a busy city street is probably a “30” sign, the result of and mischievous kids.
These examples highlight another challenge facing autonomous development: Perception is more than just seeing and sensing. It also requires context, judgment, and reasoning.
Where do humans learn path-planning? In crowded places. Observe any busy shopping mall or sidewalk, the CES exhibition halls, or outside a Taylor Swift concert ten minutes after the performance ends. Watch as that dense mass of people just walk by each other.
We develop a “social contract” as children under the guidance of responsible parents, adults, and teachers, learning the behavioral norms expected of considerate members of society. We know to give priority to seniors, parents with infants, wheelchair users, and the visually impaired or those with mobility issues.
Humans learn basic road and traffic awareness skills first as pedestrians, say by learning to cross the street, and later by learning to ride a bicycle. As we reach adulthood, many practice and then pass a motor vehicle driving exam, enabling us to individually take our skills and experience for path-planning onto public roads.
Some are more gifted at the task of driving than others, but all have demonstrated basic proficiency, which is recognized in the form of a driver’s license.
On roads and in public spaces, humans combine prediction and perception with path-planning to move autonomously, mostly without incident, injury, or even conscious thought. These movements are mostly predictable, but not entirely so. People overcome this ambiguity using knowledge-based reasoning to exhibit judgment under uncertainty.
But what counts as everyday judgment, reasoning, and uncertainty to us leaves robo-drivers flummoxed. As seen in San Francisco, robotaxis can neither predict, perceive nor path-plan human intentions with sufficient accuracy to competently share the road with humans.
Precision path-planning is less of a safety-critical issue for small, light, slow-moving autonomous vehicles such as delivery-bots. However, for 6,000-pound robotaxis traveling at city speeds, and more so for 80,000-pound robotrucks traveling at highway speeds, flawed path-planning presents life-threatening consequences for human road users.
While autonomous vehicles could excel at basic driving skills and rule-following, they clearly lack the necessary knowledge and expertise to successfully interact with humans and human-driven vehicles. Which is a problem for robotaxi and robotruck developers seeking to launch commercial operations on public roads.
What is it that makes public roads so challenging for robo-drivers?
Compared with almost every other mode of transport, public roads are complex, not complicated. Writing in Simply Complexity, author Neil Johnson observes:
“Complexity can be summed up by the phrase ‘Two’s company, three is a crowd.’ In other words, Complexity Science can be seen as the study of the phenomena which emerge from a collection of interacting objects.”
“At the heart of most real-world examples of Complexity, is the situation in which a collection of objects are competing for some kind of limited resource.”
Thus, Complexity Science precisely describes the interrelationships between motor vehicles, motorcycles, pedestrians, bicyclists, and other users on public roads.
Complex systems can further be characterized as open networks with variable demand. Witness the relative busyness of a stretch of road at rush hour compared with at 3 a.m. Complex systems have no central controller or “guiding hand,” exhibiting emergent properties leading to phase transitions from order, to disorder, and back to order. This happens all by itself, with no input required.
Confused? Just think of the everyday mayhem which occurs on public roads, which seems to come from nowhere, has no obvious cause, and then disappears into nowhere. That’s complexity, navigated by competent human drivers using knowledge-based reasoning. Most times the key requisite skill is nothing more than patience, as human perception also understands the concept of impermanence.
Compared with complexity, a complicated system is closed, predictable, and can be modeled or simulated with perfect reproducibility. A specific input will produce a predictable output.
An example is the mechanism of a Swiss watch. While incredibly intricate, the movements of every cog and wheel are known, precise and predictable. For instance, the Calibre 89 mechanism from Patek Philippe consists of 1,728 parts.
Transport systems typically involve very heavy objects moving at high speeds and that are either carrying humans or are moving quickly in close proximity to humans. Being closed and predictable, complicated transport modes are inherently safer than complex ones.
We can group examples of various transport modes together, according to those which are complicated and those which are complex, as follows:
- Complicated: Farming, mining, warehouse logistics, port logistics, aviation, rail, and private (closed) roads.
- Complex: Public roads.
Where have autonomous developers such as Aurora, Cruise, Embark, Gatik, Motional, TuSimple, Waymo and Zoox all focused their operations? Public roads, which are complex.
Meanwhile, where has autonomous development already found commercial success? Two examples are farming and mining, both of which are complicated.
This CES video shows Jorge Heraud, John Deere’s vice president of automation and autonomy, explaining autonomous tractors:
Simulation and brute-force training do not solve the reality of ambiguous, uncertain, or unpredictable scenarios common on public roads, and existing AI and machine learning techniques are too brittle to navigate phase transitions inherent in complex systems.
Commercial success in autonomous development therefore largely comes down to understanding the differences between complex and complicated systems. As former world heavyweight boxing champion Mike Tyson once observed “Everyone has a plan until they get punched in the mouth.”
Investors in robotaxi and robotruck developers look destined to learn this truth the hard way in the years ahead.
Additional discussion on the subject of complex and complicated systems is here.
We’ve heard every conceivable boast and promise of the benefits of autonomous technology from the driverless industry over the last five years. Robo-drivers won’t speed, drive drunk, get distracted or fall asleep. They will never get sick and will always be vigilant.
But what the machines lack is the necessary ability to successfully interact with humans on public roads. This is to be expected since robo-drivers, while touting the promise of excellent skills and rule-following, are devoid of human characteristics such as compassion, empathy, and feeling, as well as lacking in any sense of consequence or responsibility.
Robo-drivers predict, perceive and path plan entirely differently from humans, and consequently move and stop accordingly. That’s not a problem for an autonomous tractor plowing a field, an autonomous truck operating in a mine, or an autonomous forklift moving stock around an automated warehouse.
But for operation on public roads, which necessitates safely interacting with humans while competently navigating the uncertainty and unpredictability inherent in a complex system, the evidence is stacking up: Robo-drivers suck.
A personality disorder is typically defined as a way in which a person may think, feel, behave, and relate to others very differently from the average person.
Don’t expect NHTSA’s investigation into the Cruise problems in San Francisco to conclude with anything so definitive. The trouble with robo-drivers on public roads can easily be explained as a never-before seen condition: Autonomous Personality Disorder.
What did the tech industry’s best and brightest engineers develop with those tens of billions of dollars invested in robo-driving? Fans of Star Trek: The Next Generation will be familiar with the answer. They invented The Borg.
Let’s have a show of hands: Who wants to share the road with them?
The safety risks highlighted here are mitigated by NHTSA demanding that safety drivers b used in all autonomous vehicles operated on public roads. If safety were GM Cruise’s number one priority, it would have done so already.
Colin Barnden is principal analyst at Semicast Research. He can be reached at email@example.com.