Representing Constraints in Human Bimanual Manipulation for Transfer to Humanoid Robots

Bimanual manipulation is a fundamental skill humans acquire through years of
practice, enabling complex coordination for tasks ranging from daily activities
to expert-level performances. In robotics, achieving purposeful bimanual coordination
remains a challenge due to intricate temporal and spatial constraints.
However, mastering this skill is crucial for humanoid robots to enhance task efficiency
and enable effective programming by demonstration (PbD) and imitation
learning (IL).
This thesis develops methods for learning task models for bimanual manipulation
from human demonstrations. The work consists of three parts: The
first is the development of a taxonomy for bimanual manipulation based on
human motion analysis and findings from neuroscience and robotics. In the
second part, methods for recognizing the categories defined by the taxonomy
are developed and evaluated using new data sets for bimanual manipulation
in a household context. Finally, the implementation of the constraints resulting
from the respective categories in control strategies for humanoid robots is
shown.