Why is it so hard to make a truly self-driving car?
Chinese tech company Baidu announced this week that it has started mass-production of its autonomous minibus -- with the first 100 having already rolled off the production line.
But you won’t see these vehicles on suburban roads or even highways.
That’s because the 14-seater buses are limited to geo-fenced areas -- starting with shuttle services around nuclear power plants, as well as in senior communities in Japan. (You read that right: nuclear power plants).
Other possible areas include tourist spots, business parks and airports.
The common thread here is that they won’t have the freedom of the open road but instead be tied to a defined geographic area.
Why is it so hard to make an autonomous vehicle?
Studies show that on average, drivers make 160 decisions for every mile they travel. Speeding up, checking what's behind, signaling and changing lanes -- to name a few.
Making a car that can make those decisions correctly is challenging enough, but when you add unexpected events to the equation -- it can become a matter of life or death.
In some cases, the technology works. Like when this Tesla X in autonomous mode avoided a collision.
But it can also go tragically wrong.
Like when an Uber self-driving car -- which was in autonomous mode with an operator behind the wheel -- hit and killed a pedestrian in Arizona, believed to be the first time a self-driving car has killed a pedestrian.
Germany -- a leading auto manufacturer -- is already working on the legalities of how to approach the challenges posed by autonomous cars. It has come up with guidelines that say the machine must harm the fewest possible people and treat all life equally -- even if that might mean ruining a lamp post or injuring a dog running across the road.
Which brings us back to Baidu’s buses.
Their Level 4 designation means the vehicles can go into self-driving mode in specific conditions only.
“When we talk about Level 4 autonomy, it’s fully autonomous within a geofence, so within an area where we have a defined high definition map,” Ford’s chief engineer for autonomous vehicles Jackie DiMarco told Inverse.
“We look at autonomy as growing within a certain geofence and then expanding on there as the technology comes along, as our learning comes along and as our ability to solve more and more problems comes along.”
A little while yet
While it's possible to find autonomous vehicle use cases today -- we're still a while away from mass-market adoption: A recent poll of 300 industry insiders at a Bloomberg event found that the majority believe we won’t see Level 5 cars (no human intervention required) until at least 2030.
There are the technical challenges -- like reacting to a police car speeding past, or maneuvering more carefully on a busy road. These would require advanced AI capabilities, such as the ability to recognize objects, or predicting the actions of pedestrians and cyclists.
But there’s still another challenge that needs to be solved before self-driving cars can reach the masses: cost.
It turns out that putting autonomous driving functions into a car isn’t cheap: Current systems can add up to US$100,000 to the cost of each vehicle -- money that many consumers probably won’t be willing to spend.