An Uber vehicle in autonomous mode (with a driver at the wheel) has hit and killed a pedestrian in Tempe, Arizona. Despite that it looks very much like Uber’s vehicle was not at fault in the fatal accident, the fact that it happened at all highlights just how far the technology still needs to go to meet its objectives, especially in regards to AI.
Tempe police have stated that the pedestrian unexpectedly walked into the traffic lane, was not on a pedestrian crossing and did so from an area of shadow. In all likelihood, this means that Uber was not at fault in this incident, but it is in indicator of just how difficult it is to get even the basics right.
RFM sees three main benefits to vehicles driving themselves.
- Safety: In 2017 there were 40,100 road deaths in US, where almost all of them were due to human error. This is one of the main reasons given to remove humans from piloting vehicles – when working properly, machines should be able to do it much more safely.
- Independence: Autonomous vehicles will provide those that cannot drive themselves independence resulting in better quality of life for all concerned.
- Traffic management: By working together in ways that humans never can, machines should be able to make more efficient use of road infrastructure. This would provide shorter journey times, even with many more cars on the roads and reduce the need to build more roads.
Unfortunately, what this incident indicates is that the state of autonomous driving (and especially Uber) is very far from where it needs to be to become market ready. Furthermore, this incident is also likely to increase regulatory pressure on testing, further slowing down progress.
I suspect that the Uber vehicle’s sensors were able to detect this pedestrian, as the vehicle has top mounted Lidar capable of capturing many thousands of data points per second as well both all-round camera and radar coverage. Hence, I think that the problem was that neither the machine nor the human were able to work out that there was a pedestrian about to walk into their path. It does not appear reasonable to have expected a human driver to detect the pedestrian in this case – but this is exactly what the machines are supposed to be able to deal with.
Therefore, it is at this point that the issue of AI comes in.
If the vehicle is programmed to stop every time it detects any movement in its immediate environment, then the vehicle would barely ever move. Consequently, the system needs to able to distinguish between the movement of leaves or litter caused by wind, for example, and a potential hazard. In my opinion, this is an AI problem and is one that will prove fiendishly difficult to solve.
Predictive AIs are a product of a historical data set and the quality of those predictions largely depends on the data set remaining stable. I suspect that this incident had parameters that the Uber algorithm had not “seen” before, and as a result it was unable to correctly identify the hazard resulting in the fatality that occurred.
This is a problem that has to be overcome if autonomous driving is ever to become a commercial reality, and at the moment the playing field is not level. At the head of the pack is still Waymo, which I suspect would have had a better chance of avoiding this incident simply because it has much more experience than anyone else, and may have “seen” something like this before.
This incident is likely to slow the development of autonomous driving down somewhat, but I think that I have already taken this into consideration in my estimates. I still see liability and insurance as the biggest block to autonomous driving, as super-safe vehicles will take a very heavy toll on insurance premiums.
Hence, there is a vested interest to protect – meaning that the insurance industry is likely to aggressively fight autonomous driving, just like the taxi industry is fighting ride hailing.
I still do not expect to be renting an autonomous vehicle much before my retirement. (2028 is RFM’s prediction for commercial autonomous passenger vehicles).
This article was originally published at RadioFreeMobile