How much safer will self-driving cars be than traditional ones? They’re not on the road yet, but the Insurance Institute for Highway Safety (IIHS) believes they might only reduce crashes by about a third unless designers prioritize safety over other considerations.
According to IIHS researchers, there are basically five types of errors that lead to avoidable car crashes:
- Sensing and perceiving errors: Where the driver doesn’t see the hazard due to distraction, impaired visibility or simply not recognizing it
- Prediction errors: Where the driver misjudges, incorrectly estimates or makes incorrect assumptions about how other vehicles will behave
- Planning and deciding errors: Where the driver speeds, drives aggressively, follows too closely or makes other decisions that don’t prioritize safety
- Execution and performance errors: Where the driver mishandles the vehicle, such as overcompensating or failing to take the correct evasive maneuvers
- Incapacitation: Where the driver falls asleep at the wheel, has a medical problem or is impaired by alcohol or drug use
Would a fully autonomous future have no car accidents?
The researchers examined police records for over 5,000 crashes from the National Highway Traffic Safety Administration’s National Motor Vehicle Crash Causation Survey involving two factors: at least one of the vehicles was towed, and emergency medical responders were called to the scene. This was a nationally representative sample.
They then imagined a world in which all vehicles were fully autonomous. The researchers determined that the fully autonomous vehicles would be able to completely avoid crashes caused by incapacitation or sensing and perceiving errors, assuming that the vehicles’ sensors and cameras were operating properly. Together, these types of errors accounted for 34% of all of the crashes.
However, the other types of errors — prediction, planning and deciding, and execution and performance — involve behaviors that require some degree of judgment. That makes it more difficult to program a vehicle to handle these types of tasks.
The assumptions of that programming are essential. The vehicles’ designers need to get the vehicle to prioritize safety over other considerations such as convenience and speed. To be any safer than a human-driven vehicle, the system would have to prioritize safety to a greater degree than human drivers typically do.
To be safe, fully autonomous vehicles will need to do more than follow traffic laws. They will need to respond to changing conditions and to human behavior that can be hard to predict. They will have to exercise good judgment in those conditions.
It’s important to note that there are no fully autonomous vehicles on the road today. Some vehicles can drive with little input from the occupant of the driver’s seat, but all of them require constant vigilance. Never assume that your “autopilot” function allows you to take a break from watching the road or directing the vehicle.