Automation of construction machinery has the potential to improve efficiency and safety during the construction process. However, most construction machinery is directly teleoperated limiting the control to a single paradigm. Moreover, the abundance of non-linearities due to the design of pumps, valves and the interaction of different actuators complicates planning precise movements. Complicated motion planning optimization can be applied to program the movement in the desired manner which often requires a model representing the system dynamics. Whereas this approach is promising, modeling the system dynamics is a formidable task. In this work, we present a framework that uses a probabilistic approach to learn movements from expert demonstrations. In this way, efficient movements can be learned without explicitly estimating system dynamics. Here, efficiency is defined by the experiences of the human operator which makes the programming able to benefit from existing knowledge of operators. The performance of the proposed scheme is evaluated with a real demolition machine BROKK 170.