Tasks such as non-prehensile manipulation (e.g. pushing) pose
difficult computational challenges for robots.  It is difficult to
develop a reactive controller capable of completing these tasks
within complex and unpredictable environments.  In contrast, humans
typically complete these tasks repeatably and with ease.  The goal
of this paper is to describe an approach we have been developing to
incorporate human input via a shared autonomy interface to improve
solver performance (e.g. convergence).  We provide early empirical
evidence that our approach leads to higher success rates for planned
motions, and fewer iterations for the solver.