For more options, greater safety and cheaper prices autonomous cars could use plug-n-play AI. Obviously I am not an expert on driving cars, but this is how I might do it.
Some Ideas
For maximum safety, two independent AIs could compare notes. Each car could learn more from being used by its driver in its specific location. The can could take verbal and entered directions perhaps.
It is important that driving cars become available for all, like Ford's model T, and not an expensive feature unrealistic for most to buy.
So first I would create a standard where most driving cars could change out each hardware-software components. Just unplug it and put a new one in. It could be like a box, with computer chips, software, memory, etc.
Universal Language
The universal language could be a 3D space described in physics. A mapping of the car's world to mathematics and other data. This would be the first job of an AI. The second would be decision making based on this data.
So now imagine a car and the AI takes all available data and creates a world, guessing some facts such as the weight of a nearby car and knowing others such as distance and speed. Any AI following standards, could analyze this data and make its own decisions.
In this 3D mapped space could also be data on the positions of cameras, lasers, sonar, radar and a summary of what each is detecting.
Plug and Play AI
This is a pretty straight-forward idea, and you can skip this section.
This is what I image. There would be places for 5 or 6 computer boxes. The first box would perhaps come with the car or be custom to it. It would take the data from the devices and make its own best guess of its 3D world. (It might be split into three, where the first interprets using chips mega amounts of data, if this is not done somewhere else, and the second creates the 3D world, and a third could add statistical guesses such as the weight of nearby minivan, with its people inside.)
The second box would contain computer AI for interpreting the information and making decisions. It could be split into three as well. One that communicates with the car's human-interaction devices, one that maps out a course, and one that drives the car.
A third place could be available for a backup AI. It could accept and reinterpret the 3D world and data from box 1 and 2. For example, it might guess a different weight for a vehicle or predict different reactions from nearby drivers. This third AI, could judge slicker ice on the road, assign more weight to a vehicle or guess a different path be taken by a driver. Then it would make its own decision. This could be backup if box 2 fails or feels confused. But its more important role might be how to handle a dangerous situation. It might just agree with box 2 on every decision while no faulty decision making is detected.
A fourth independent box could be an AI specializing in arbitration. If backup AI completely disagrees with the first AI's choice of action, it could protest to the arbitration AI, which maybe could override and protect the vehicle from accident.
Box 1 which gathers information, should have its own fail-safes and maybe even be judge 3D space without all sensor equipment working. There can be camera inside, outside and under the car and various other sensor types. I believe however that technologies such as radar should only be used carefully for human safety reasons.
To Simplify
All of a sudden this all seems really complicated. Basically, it comes down to standards and options. If there is competition on driving AI then it brings prices down and open source software may also compete.
There is some complexity to how cameras and other gear work together to understand a situation, but this could also come down to smart software, able to judge the 3D world around it, using any combination of locations and types of sensor gear.
The idea of multiple AIs comparing notes and voting on decisions of safety may be unnecessary, but combining independent thinking may add depth to decisions.