The emergence of AI from 1990 to 2010

The starting point of the history lesson in technology is the year 1990 which was the end point of the 4th computer generation. Around the year 1990 most modern computing technology like 32bit CPU, the C++ programming language and Internet protocols like TCP/IP were invented. Also videocompression and many video games were available.

Artificial Intelligence and robotics belongs to the 5th computer generation which has emerged after the year 1990. The starting point are heuristics algorithms. In contrast to a normal algorithm which can be search in a sorted database, a heuristic algorithms is using external knowledge to fasten up the performance. Entry level heuristics algorithms are based on an evaluation function. The most famous one is A*. In contrast to a simple traversal of the state space, A* is using ja distance function to follow only nodes in the graph which are near to the goal. This allows to search much faster.

Modern AI technology including robotics is build up on heuristic algorithms. Its not possible to solve AI somehow else because problems like motion planning and grasp planning are so complex, that always sort sort of heuristics is needed to fasten up the processing speed. The only debate is about which sort of heuristics fits well to a certain problem.

Early and easy to explain heuristic algorithm are based on the previously mentioned evaluation function which can be used to implement path planning algorithm and chess playing computer programs. In computer chess, the evaluation function determines a score for the current board, this score is used to prioritize the search in the state space. Such an algorithm can be executed on normal desktop computers which were available in the 1990s like a 286'er PC. An evaulation function alone can't solve more complex robot problems. The problem is that for high level tasks like "grasping an object" the evaluation function is unknown. These obstacles has prevented, that robotics was available in the 1990s so there was a longer process needed to heuristic algorithms.

A promising approach developed during the 2000s in parallel to the Deep learning boom was called feature engineering. Feature engineering tries to create evaluation functions for any domain. The question is which statistical variables are important for a domain. For exampole a robot hand might have a touch sensor, while a car might have the features speed and direction. The principle of features allows to create more advanced evaluation function which can solve lots of robot problems.

A third and most advanced tool to create modern heuristic algorithms is natural language. Instead of treating robot control problems as purely mathematical subjects in a multidimensional numerical space, the idea is to communicate with a robot over a language interface similare to a text based adventure game. The ability to introduce grounded language into existing heuristi algorithm porovides a powerful framework to create advanced robot systems since  the year 2010.