Tesla FSD Training — The Rules-Based Automation That’s Needed

Europe is in many aspects just a little bit different from the USA. One of the things that is different is the traffic laws. They have become more harmonized across Europe in recent years, but still not identical. Crossing a border implies getting into a jurisdiction with a different set of rules.

A simple example is the right of way when entering an “Autobahn” or “autosnelweg.” The sign is the same blue rectangle with two white parallel vertical lines crossed by a horizontal bridging line, but on the German Autobahn, the vehicle on the on-ramp has right-of-way to enter the highway. On a Dutch autosnelweg, the vehicle on the on-ramp has to cede right-of-way to all the traffic on the highway. It has to stop at the end of the on-ramp if there’s no safe way to enter. Not knowing this subtle but essential difference when crossing the border, there can be a lot of dangerous traffic violations.

In my previous article about Tesla FSD (Full Self Driving) training, it was mentioned that not only is all driving rule based, but also that breaking the rules is rule based. The rules of breaking the rules are informal, but very important.

For example, when encountering a speed sign that mandates lowering your speed, the Dutch rule is to lower the speed after the sign; in Germany, the rule is to lower the speed before the sign. Accelerating is the same, in reverse. The Dutch will accelerate when seeing the sign in the distance, the Germans will accelerate after the sign.

In Switzerland, ascending traffic has right-of-way over descending traffic. The Dutch have no such rule, there are no mountains. To safely drive after crossing the border, the new set of rules have to be implemented. This should be easy for an AI that drives rules based. Training the AI ​​with just driving examples, without making clear that one set is for one country and the other set is for a different country, however, things can get difficult.

A more fundamental difference between European and American driving is the “keep your lane” versus “keep to the right” regulation. It is one of the fundamental regulations that has a large impact on driving behaviour. Overtaking on the right is a very serious offense, it can cost you your driver’s license.

Some may think this article is attacking the neural net approach of Tesla. That is far from the truth. In the previous article, system architecture was mentioned. Important parts of the system architectures are the vector space. I do not know whether the vector space is moving around the (static) car, or the car is moving through a static vector space. It is a subtle difference. From afar, it looks like the car moving through a static vector space is preferable, but practical implementation can favor the other option. What is important is that the objects in the vector space are persistent, even when they are out of view.

Classifying all objects in the vector space, buildings are static, humans are dynamic, parked vehicles are likely to stay static, driving vehicles are dynamic, and a thousand other objects have to be classified. That is something only an AI can do. All objects that are not static have to be evaluated on their likely behavior. Also a very hard task for the AI.

All road signs, both static and from dynamic (human) sources must be interpreted. The AI ​​must be able to read (many languages) to understand the qualifiers, like “when raining,” “when low visibility,” “when bad visibility,” etc.

After all those evaluations, interpretations, guesses about expected behavior of other traffic participants, the rule-based automaton must decide on the actions to take and how to implement them.

It is not best practices from millions of examples of lousy drivers (excuse me, I am not anti-American, but statistics paint a very clear picture), it is optimally implementing the rules. A normal driver makes about one mistake every minute. In heavy traffic, more often. At least 90%, probably over 98% of all examples Tesla has collected, are examples how not to drive. And that one mistake per minute is from European driving.

Over a year ago, I wrote an article about finding the best trainers/testers. It is now more applicable than ever. The development of FSD is not two steps forward, one step backward like in some processes. There is no known reason to do this in developing a software system.

The AI ​​is for higher level recognition and decision making. The rules are for a perfect execution. Going left at a traffic light should be a piece of cake, not the gamble it is now. To be clear, the AI ​​should indeed select the rules that apply. The AI ​​should supervise perfect execution of the rules. When there is a speed sign, the AI ​​should select the “adjust-speed-to-speed-sign” rule applicable to the location. When there is traffic on an on-ramp, the AI ​​should select the appropriate rule. When driving in the mountains, use the mountain driving rules. Rules are simple, stupid, easy to follow, and very important to drive in the best way possible.

The advantage of a rules-based system is the ease with which a new country can be added to the FSD system. The driving intelligence stays the same, only the rules are different.

For those who think me too harsh or negative, my driving capacities are not getting better. At my age, I need a FSD system in my car yesterday.



Appreciate CleanTechnica’s originality? Consider becoming a CleanTechnica Member, Supporter, Technician, or Ambassador — or a patron on Patreon.


Have a tip for CleanTechnica, want to advertise, or want to suggest a guest for our CleanTech Talk podcast? Contact us here.

Leave a Comment