The feasibility of autonomous vehicles for general use in public traffic has long been questioned (see “Artificial intelligence faces serious roadblocks,” Asia Times, July 18, 2017). Here we are, years later – after billions of dollars of investment – with the reality that the technology still does not meet the requirements for reliable use in general traffic.
This deficiency was highlighted last month when the California Department of Motor Vehicles suspended operations of General Motors’ autonomous-driving subsidiary, Cruise. The suspension was triggered by an incident on October 2 when a human-driven vehicle struck a female pedestrian, propelling her into the path of a driverless Cruise car.
The driverless car collided with her, came to a halt, and then attempted to pull over to the side of the road, dragging her about 6 meters. Evidently, the car’s computer program failed to meet the demands of this situation.
It’s not due to a lack of technological investment – billions of dollars have been poured into creating autonomous vehicles equipped with sophisticated sensors and refined software through extensive road operation.
There was a belief that these vehicles were becoming equipped to handle the intricacies of general road traffic. In addition to GM’s setback, TuSimple Holdings, a self-driving-truck maker, is liquidating, causing billions of dollars of market value to evaporate.
Despite these periodic setbacks, optimists remain convinced that artificial intelligence (AI) software will eventually enable vehicles with superior driving performance, saving countless lives by minimizing accidents caused by human drivers.
Prior to the suspension, GM had projected US$50 billion in revenues from its driverless-taxi operation by 2030. Reflecting this optimistic forecast, the subsidiary was valued at $30 billion in a 2021 funding round.
The wonders achieved with new AI software technology prompt the question: Why is this application so challenging?
These self-driving vehicles are unique robots, with a key distinction from familiar robots designed for industrial or commercial environments. Conventional robots are designed for programmed functions anticipated in their application, such as welding or painting in manufacturing.
If unanticipated events disrupt the assigned process, the robot stops or engages in specified procedures to minimize damage to the production line. In typical production environments, these default-programmed procedures can be implemented with confidence.
Now, consider an autonomous vehicle. If designed to operate in a controlled environment where events are predictable, such as a closed street or limited track, robotic performance can be programmed because the anticipated events are limited. For instance, inter-airport rail lines have driverless trains that operate safely because track access is controlled. Should an object obstruct the train, the sensors will detect it and stop the train.
However, an autonomous vehicle designed to navigate freely in open traffic, like the GM vehicle, needs to operate properly under practically unlimited conditions. In effect, the vehicle is expected to emulate human intelligence. Despite claims that such vehicles will outperform human drivers, human intelligence remains key because of its unique capacities.
Traffic accidents are, by nature, random events – each different in some elements. Dealing with such situations involves rapid thinking and accessing information not obvious to the vehicle robot, such as peripheral information that automated sensors may not capture. No amount of training can prepare the vehicle to respond appropriately to unlimited traffic conditions.
Millions of kilometers of traffic driving have been used to “train” these vehicles, and billions of dollars have been invested in software improvements. Nevertheless, they periodically encounter situations where the “trained” conditions do not match reality, and the vehicles perform poorly – worse than expected from human drivers.
The questions remain: Will further investment overcome these robotic limitations, and will autonomous vehicles outperform human drivers? My guess is that they will find their niche in controlled environments. Replacing general human driving? Hmmm.
Henry Kressel is an inventor, technologist and author, as well as a long-term private equity investor in technology companies.