In order for a robot to carry out meaningful tasks in diverse and complex en-
vironments, it needs the ability to autonomously discover salient parts of its
environment that are relevant for successfully carrying out its tasks. One sub-
set of such tasks includes those that involve manipulation of everyday objects.
Autonomously acquiring a task relevant model of a previously unseen object
would expand the set objects that a robot can eectively (correctly and e-
ciently) manipulate. Such a model needs to be as simple as possible to decrease
the overhead of acquiring it and to increase the likelihood of it transferring
to similar objects. A hypothesis is that the kinematics model of an object is
enough to allow the robot to eectively use this object. For this project, we
will assume that we have some robot that can eectively manipulate a previ-
ously unseen object to obtain sucient information from its sensors to infer a
complete kinematics model of the object. Furthermore, the only sensor that we
will use is a stereo camera. Thus, our goal becomes to discover the kinematics
model of previously unseen objects using only vision.Discovering Objects' Kinematic Structures