Whole-body control in unknown environments is
challenging: Unforeseen contacts with obstacles can lead to
poor tracking performance and potential physical damages of
the robot. Hence, a whole-body control approach for future
humanoid robots in (partially) unknown environments needs
to take contact sensing into account, e.g., by means of artificial
skin. However, translating contacts from skin measurements
into physically well-understood quantities can be problematic
as the exact position and strength of the contact needs to be
converted into torques. In this paper, we suggest an alternative
approach that directly learns the mapping from both skin
and the joint state to torques. We propose to learn such
an inverse dynamics models with contacts using a mixtureof-contacts
approach that exploits the linear superimposition
of contact forces. The learned model can, making use of
uncalibrated tactile sensors, accurately predict the torques
needed to compensate for the contact. As a result, tracking of
trajectories with obstacles and tactile contact can be executed
more accurately. We demonstrate on the humanoid robot iCub
that our approach improve the tracking error in presence of
dynamic contacts.