The ability to perform human-like manipulation actions using artificial robots is a major requirement in service robotics.
This is a problem related to both high-level symbolic reasoning and low-level control systems.
This work proposes a multi-layer framework to fully define and execute a wide range of such actions
in a generic and generalizable fashion.
We present the details of action definition and execution and collect them into a re-usable software library.
The first contribution of this thesis is definition of high-level and low-level components of actions as well as a clear mechanism to link them in execution.
To demonstrate the ability of execution system to generalize on a wide range of actions and objects, a large set of 300 trials is performed.
The success rate of each action is calculated and the failure cases are analyzed.
The second contribution is applying the concept of structural bootstrapping to get action parameters from human demonstrations and previous experiences.
Here, several human demonstrations obtained by different methods are processed.
New instructions are executed based on previous knowledge which enables system to go beyond hard-coded actions.
Last contribution is to integrate the actions with a symbolic decision making framework to benefit from the advantages of the state-of-the-art in planning.
Here we deal with grounding symbolic operators of planner to solve complex tasks such as making a simple cucumber salad.
We also feedback the faults of execution to the decision-making system which enables learning new operators through a human operator.
Keywords: Library of Actions, Manipulation, Semantic Event Chain
