Not necessarily, according to a recent study. Their ability to out-perform humans in terms of safety will depend not only on the accuracy and reliability of their sensors but also on how they weigh the importance of safety.
The Insurance Institute for Highway Safety (IIHS) studied a set of nationally representative auto crashes and examined their causes. Then, the group tried to predict whether an automated vehicle might have been able to avoid the crashes.
The researchers divided the crashes by cause into five categories:
- Sensing and perceiving errors: The driver was distracted, unable to see or failed to recognize a hazard before it was too late
- Prediction errors: The driver wrongly estimated another driver’s speed, misjudged a gap in the traffic, or incorrectly assumed another driver would be have in a certain way
- Planning and deciding errors: The driver didn’t follow traffic laws, was aggressive, followed too closely, etc.
- Performance and execution errors: The driver made mistakes controlling the vehicle, such as overcompensating or failing to make the correct evasive maneuver
- Driver incapacitation: The driver was impaired by substances, had a medical problem that interfered with their driving, or fell asleep
In some cases, the researchers determined that a crash was unavoidable due to vehicle failure.
Would a self-driving car do better than the human drivers had done? The researchers decided that automated vehicles could be expected to do better in at least two categories of driving errors: sensing and perceiving and driver incapacitation.
To be sure, not all sensors are perfect, and some automated vehicles would suffer from vehicle failure. Indeed, the researchers pointed to one semi-automated vehicle that failed to detect a pedestrian, failed to predict she would cross the road, and failed to take evasive maneuvers when she did. She was killed.
Sensing/perceiving errors and driver incapacitation accounted for about 34% of all the crashes. In order for the automated vehicles to prevent the other two-thirds of crashes, they would have to be designed not only to make no performance and execution errors, but also to avoid prediction, planning and deciding errors.
Whether they could do that will depend in large part on how the cars are designed and programmed. They would need to be able to recognize potential hazards, such as pedestrians, bikers and human-driven vehicles, that might not behave as predicted and compensate quickly when the situation changes.
Prediction, planning and deciding all require judgment. Can engineers design and program vehicles to have better judgment than humans?