Extending actuator-level control for robotic systems using distributed dynamics computation and incremental learning

Actuators are fundamental components in robotic systems such as manipulator arms, since
they enable a system to actively and physically interact with its environment. In classical
industrial production environments, the focus of the optimization of this interaction
has been precision and speed. The individual operating steps are typically tailored to a
certain stage of the production and are hard programmed. Further interactions – including
unintentional – are not considered within this scope and are therefore prevented by strict
separation of workspaces for safety reasons. However, such limitations cannot be enforced
if robotic systems are also to be used outside such defined workspaces. This is inevitably the
case when they directly support or assist human individuals and thus share their workspace
with humans, as is increasingly the case in industrial assembly processes, the service sector,
and for assistance in the field of rehabilitation and household. In such cases, rigid position
control of the joints is not sufficient, as it regards any deviations, irrespective of the current
situation, as an error and increases the actuation forces in response. It does not take
into account what driving forces are actually reasonable and necessary, nor whether a
new situation has arisen that is causing the deviations. The aim of this dissertation is to
extend the capabilities of the actuator-level control so that it incorporates the relationships
between the generated forces and torques with the motion and external forces and learns to
recognize new situations from its own sensory data. For this purpose, this work combines
three approaches that build on each other and examines them experimentally. Firstly,
methods and software tools are being developed that facilitate the creation of dynamic
models based on physical insights in such a way that the missing parameter information
can be obtained more easily from experimental data. The experimentally determined
models in combination with an advanced motor control form the basis for safe compliant
motion control of a manipulator arm without additional external sensors. Secondly, based
on this, a method is developed that incorporates the dynamics locally at the actuator level
by means of a distributed computation. This makes it possible to design more modular
robot systems and to relieve a central computer, as well as to reduce the dependency on
the data communication and associated latency that would normally be necessary for a
central calculation. Thirdly, based on the concept of Adaptive Resonance Theory, a kind of
episodic memory for recognizing different situations is developed. To achieve this, sensor
data from robotic actuators are pre-processed in frequency domain and continuously –
incrementally – learnt by an artificial neural network. It is shown experimentally that
a robot system can distinguish collisions during an arm movement or jamming of parts
during an assembly task from previous undisturbed executions.