Manipulation planning using model-based belief dynamics

Planning in partially-observable domains require an agent to fuse prior knowledge with observations to update belief and to search for optimal plans that reduce uncertainty with respect to the task. This requires a knowledge organization that captures the underlying dynamics of the belief space and its probabilistic dependency on actions. In this paper, we present a functional representation for organizing knowledge about the environment in terms of interaction statistics. The representation utilizes a uniform, domain-general description of state that applies to a wide variety of tasks. We show how a planning algorithm can exploit this knowledge representation to build plans directly in the space of control actions. Given incomplete state information, the planner interacts with the task to acquire the information required to solve it. We illustrate the approach in multiple demonstrations of an object recognition task.