Traditional industrial robot arms are widely used in today's manufacturing processes. They are very precise and versatile. However, due to their weight, their use cases are very limited in mobile applications, aerospace or human-machine interaction. The development of lightweight manipulators is therefore necessary in these fields of application. Not only are they easier to transport, but as a result of their lower inertia, faster movement is also possible. This comes with consequences though. The robot arm is far more flexible and vibrations are induced by every movement. Therefore, a suitable control algorithm needs to damp and prevent them if possible.
In this thesis a two joint robot arm mounted on a flexible base will be controlled using a reduced SpiNNaker2 prototype. With its 16 instead of 152 ARM Cortex M4 processors and hardware accelerators, it still provides massive parallel execution and fast neural network computation capabilities. With Reinforcement Learning a control algorithm was found that not only seams to be suitable for the control problem at hand, but also tests the performance of the hardware in this regard.
This thesis will guide through the implementation of Reinforcement Learning for such a manipulator configuration including the creation of a simulation environment, the training procedure as well as a heavily parallelized realisation on the SpiNNaker platform. For comparison an input shaping approach is applied as well.