Abstract
Motion planners for contact-rich manipulation are limited by current implementations of contact samplers. In this report we present initial progress towards developing a generalizable and performant goal-conditioned contact sampler. We explore the use of randomized smoothing and analytic smoothing to overcome the flatness of the cost landscape under exact dynamics, thereby enabling the use of gradient based methods to improve sampled configurations. We also show the effects of different smoothing and gradient descent parameters on the cost landscape and gradient descent trajectories on simple planar pushing examples.