Autonomous Autos – Sleight of hand

Baidu’s vehicles are not really driverless.

• Baidu, like Waymo and Cruise, has received a license to operate “driverless” vehicles on Chinese roads but the reality is that just because there is no one in the driver’s seat does not mean that there is no driver.
• Baidu now has permission to operate vehicles with no humans in the vehicle in Wuhan and Chongqing but only during the daytime and only within strictly defined parts of the city.
• Furthermore, it looks like riders can only get picked up and dropped off at specific places as opposed to regular taxis which can pick and drop off from anywhere.
• Baidu is referring to these as “driverless” but I am pretty sure that there are drivers it is just that they are not sitting in the vehicle.
• Instead, they are monitoring the vehicles remotely and are ready to intervene if necessary.
• This is how autonomous driving companies can give the illusion of a driverless service, but the reality is that they are probably more expensive to operate than having human drivers in regular vehicles.
• The promise of robotaxis is not the novelty of having a car drive itself but the economics which promise to reduce the cost of ride-hailing from around $2 per mile to around $0.30.
• It is at this point that the proposition becomes compelling but we are very far from that today.
• This is mainly because, despite the hype and the marketing, machines are still really bad at driving vehicles on roads populated with other humans and changing conditions.
• RFM Research has previously identified that this is largely due to problems with machine vision where the system makes a mistake in interpreting its surroundings and then makes a bad decision based on the erroneous interpretation.
• Machine vision systems are deep learning based which means that they perform well on tasks where the dataset is finite and stable.
• The open road is neither of these things which is why autonomous driving is getting constantly delayed and why the services that are available are so limited in scope.
• Rigid geofencing, limited pick-up and drop-off points and the choice of location can greatly reduce the instability of the data set meaning that the error rate of the system drops meaning that it can perform to a much better standard.
• However, these limitations greatly reduce the utility of the service and will need to be removed before the service becomes useful.
• Furthermore, humans are still needed to oversee the vehicles, but this is where some productivity gains can be made.
• Early on, these vehicles will require one human per vehicle but as the vehicles get better this should be able to be reduced.
• When one human overseas multiple vehicles, then the economics start to improve but it will not be until there are no overseers that the proposition will really gain traction.
• Hence, I think that there is still a very long way to go and big announcements about cars with no humans are a red herring as the humans are still watching just not from within the vehicle.
• RFM’s target of real commercial autonomy in 2028 was set in 2017 and still looks good today.