There is an increasing demand of robots able to assist people in
everyday home tasks. Some of these tasks, specially garment-related
tasks (e.g. laundry, ironing, etc), are currently a big challenge for
robots due to deformations and the high number of possible
configurations the garment can adopt. Ironing, additionally, requires
manipulation and control of the iron, as if performed incorrectly it can
create even more wrinkles. In this work we present a method for robotic
ironing with an unmodified iron and ironing board and a full-body
humanoid robot, inspired in the way people perform the ironing task. In
this method, the robot performs an ironing action controlled in velocity
with feedback from a force/torque sensor. This action is later analyzed
by the robot using computer vision to determine if any wrinkle exists,
either pre-existent or created by the ironing action. The 3D vision
algorithm segments the garment surface to be ironed and computes a
Wrinkleness Local Descriptor (WiLD) that determines the location of the
wrinkles on the garment. Wrinkles are processed using computer vision
techniques on an flattened image created from the WiLD descriptors
computed in the prior stage, resulting, if present, in an optimal
ironing path that reduces wrinkleness and avoids creating new wrinkles
in garments when ironing. The experimental results show that using our
method the humanoid robot is able to successfully iron several garments
with results similar to the expected from a human performing the same
task.