For robots to operate in unstructured environments, they are required to interact with objects through contact. Those contacts may be used to push objects to the side, deform objects, or manipulate objects in-hand. This thesis addresses the problem of controlling robots to exploit contacts to manipulate objects. Being able to anticipate the outcome of such physical interactions is essential for robots to gain true autonomy. However, contact interactions are particularly challenging to reason over in model-based control approaches due to the discontinuous nature of contacts. Moreover, interacting with objects the robot has not interacted with before will naturally lead to uncertainty in the prediction of contact dynamics. For instance, the robot can not anticipate the mass distribution of an object before making contact, which requires the robot to reason over possible outcomes before touching the object in a potentially unfavorable or unsafe way.

Throughout this thesis, we formulate the problem of contact-rich manipulation with a robot manipulator as a model-predictive control problem. We explore stochastic optimization to plan for robot control trajectories in realtime. We show that the stochasticity in the optimization process enables the algorithm to explore the space of contacts without relying on local gradients or discretization of the contact space. We furthermore study how uncertainties in the physical properties of the object propagate through the contact dynamics and how the robot can actively reduce such uncertainty by exploiting favorable contact modes and sequences. We integrate the above contributions into a planning and control framework for robots to manipulate objects through contacts in realtime. The framework is evaluated in a series of robot experiments, demonstrating robots autonomously performing dynamic hand-overs, push objects to a moving target, play air hockey, and manipulate objects robustly using two arms in the presence of uncertainty in the dynamics of the object.