The transiton from closed to open robotics systems

The last AI winter went until the late 1990s. In this period, some robotics were built by the engineers and some AI algorithms were designed but all of them failed. The only thing working reliably was a simple CNC machines which were used in a static factory setups to cut a piece of metal. Even a simple pick&place robot for an assembly line was beyond the capabilities of the 1990s technology.

Today's robotics in the 2020s is much more powerful and this improvement can be explained with a paradigm shift. Robotics until the 1990s was organized with a closed system assumption. the idea was to treat a robot as as a microcontroller which runs a software in the batch mode. It was a mathematical and a computer science artifact which was controlled by deterministic algorithms implemented in a programming language like C/C++. The assumption in the 1990s was, that such a paradigm is powerful enough to create artificial intelligence. The assumption was that the existing tools like a 16bit microcontroller, a PID controller, a Kalman filter or a C compiler allows to build robots.

What the engenners didn't know was that the mentioned tools are equal to a dead end. Even with today's knowledge its not possible to build a robot with such an equipment. What is needed are different tools located outside of computer science which allows to build open systems. these advanced tools are: 

- motion capture: a human actor demonstrates a movement for a camera
- grounded language, a vocabulary to communicate with a robot
- a multimodal dataset which stores mocap data and semantic annotation in a database

These tools were missing in the late 1990s. Not because of technical constraints, but because of missing understanding for the difference between open and closed systems. A robot can be built only by one of the principles: either the robot understands natural language or it doesn't. Either the robot can playback motion capture data or it can't.

The dominant reason why these advanced tools were missing in the 1990s is because they are located outside of mathematics and computer science. Motion capture has its root in biomechanics and in animated movies. It was introduced for Rotoscoping which allows to draw cartoons. While grounded language has its root in linguistics which is located in the humanities which is the opposite of mathematics. 

In the 2020s computer science has redefined its own boundaries because the former restriction to mathematics and algorithm theory was not able to solve robotics problems. No matter which mathematical theory was applied to robot control, all of them failed. The dominant problem in robotics control is the state space explosion. A robot has many degree of freedoms and planning inside the error map of such a kinematics chain will need too much CPU cycle. There is no algorithm available which can search faster in the state space, but the mathematical perspective itself is the obstacle.

The inner working of a state of the art robot from the 2020s can be explained as a machine who understands English commands and has access to a motion capture database. These tools combined allows the robot to solve complex problems like biped walking and grasping objects. From an AI perspective, the intelligence of the robot isn't encoded in a computer program but the intelligence has its origin outside of the robot, namely motion capture data and verbal commands. The robot is reduced to a minimal device which executes an existing trajectory with the servo motor and is converting a command into action. For example, the human operator may say "move with trajectory #12", after fetching the trajectory from the database the robot activates its servo motors. Strictly spoken the intelligence has its origin not in the robot but the intelligence comes from the environment namely the human operator.

Robots constructed as open systems can be seen as communication devices instead of computing devices. They are not running a program similar to a Turing machine but they parsing a message similar to a Telefax machine.