The 2 DoF SCARA-robot arm PASCAL (Pneumatically Actuated SCARA Controlled by Adaptive Learning) is driven by two standard industrial pneumatic cylinders (of 63 and 32 mm diameter) with piston rod. They are actuated via proportional valves by 6 bar air supply. This robot arm can be operated either horizontally or vertically (shown in the figure), thus allowing for different control tasks and for scaling of the control problems from a single axis without gravity to two strongly interacting axes with the same set-up.

This pneumatic drive system features a standard industrial rodless cylinder (16 mm diameter) and can be loaded with different load masses. The aim is to have a simple and easy to operate, but practically relevant drive system with challenging control problems. It is used as an intuitive example to investigate different control principles and learning architectures. It is a simplified version of the PASCAL robot arm at the same time. Both set-ups thus allow investigations on how and how far control principles can be generalized and deriving more general engineering principles.

On the one hand, this pendulum cart is just another real implementation of an inverted pendulum at a first glance and acts a well known control benchmark. Though it is a very simple construction, it allows swinging-up the pendulum and is thus a suitable platform to investigate modular control principles at a more general level. But due to its simple construction, several practically relevant effects arise like e.g. slippage and backlash or limited control energy. This pendulum cart thus forms a real control challenge on the other hand.

This is standard set-up offered by Quanser. Because it has less severe control problems than the pendulum cart, this inverted pendulum acts as a starting point for controller design. Transfering the principles to the pendulum cart allows drawing more general conclusions.

A simple optical sensor consisting of 16 photodiodes is used as a source of real, i.e. noisy and uncertain input signals which are taken from ordinary game cards. These cards are mounted on a rotating table which rotates with variable speed. This set-up is used as a test platform for learning pattern recognition. In best case we were able to recognize each card by presenting it only one or two times for training of up to 2 m/s without synchronization.

The TULPEN platform (TUbe Learning Process Environment) is a real set-up with strong nonlinearities. Like our other demonstrators it is used for investigating learning control principles. Its different character enables studies on the general applicability of our methods.