Autonomous driving – Full stop ahead.

Whack-a-mole is not good enough.

• A Chinese demonstration of the inadequacies of Tesla’s autopilot also clearly shows that the current methodology of rushing out software patches after the event is simply not good enough for autonomous vehicles.
• One only has to look to Boeing and the severity of the issues it is suffering as a result of alleged bugs in its stall system for an indication of the critical nature of autonomous vehicle reliability.
• Researchers at the Tencent Security Lab managed to get the Tesla Autopilot to veer into oncoming traffic by putting tiny barely perceptible stickers onto the road.
• These stickers were perceived by the autopilot as lane markings with a high level of confidence resulting in the vehicle swerving into what would have been oncoming traffic.
• The problem with vehicles is that, in many cases, patching the error after the problem has occurred is not good enough as human lives are at stake.
• This is why the automotive industry has much higher standards of safety and reliability than most Silicon Valley products.
• This problem also highlights one of the main problems of using deep learning models to drive vehicles.
• Deep learning excels in situations where the data set is both finite and stable.
• This ensures that the model can realistically be shown practically every data point and that those data points will not change.
• However, the road is neither of these two things meaning that a deep learning algorithm on its own is very unlikely to be able to drive a vehicle to anything like a high enough safety standard.
• In the traditional Silicon Valley mindset this would mean patching the system every time it breaks, but when people start dying this is not going to be anything like good enough.
• This fits with RFM research (see here) which has highlighted the limitations of deep learning which could lead to a 3^rd AI winter.
• The net result is that a different approach to autonomous driving is needed and rethinking the system again from scratch is going to take time.
• I am sticking with my 2028 commercial availability for autonomous vehicles for the moment, but it is starting to look a little precarious.