|dc.description.abstract||State-of-the-art robotics research has been progressively focusing on autonomous robots
that can operate in unconstrained environments and interact with people. Specifically,
manipulation tasks in Ambient Assisted Living environments are complex, involving an
unknown number of parameters. Recent years show a trend of successfully applied machine learning approaches affecting day-to-day life. Similar tendencies are perceivable
in robotics, existing methods being enhanced with learning-based components.
This thesis studies approaches for incorporating task-specific knowledge into the
motion planning process that can be shared across a heterogeneous fleet of robots. A
step towards data-driven strategies will allow the field to break away from manuallytweaked, heuristics- or state-machine-based solutions and provide good scaling properties, while maintaining operation safety around humans at a very high level.
The presented work proposes a motion planning framework employing Learning
from Demonstration to encode task-specific motions, facilitating skill-transfer and improving state-of-the-art in motion planning. Resulting algorithms are compared against
other methods in a series of everyday tasks.
While different optimisation methods have different benefits, it is possible to build
them into systems that both generalise and scale well with the number of tasks and
number of robot platforms. This thesis shows that optimisation-based planners are ideal
for incorporating prior knowledge into a motion-planning system.||en