Physical robots suffer from uncertainties. These uncertainties arise from many sources, like when objects move in the environment, sensor data has noise, or when the robot cannot perfectly execute a given command. Motion planning under uncertainty requires more deliberative reasoning about future actions and observations in order to pick the best action at each state in time to achieve a task. This mapping of states to actions is known as a policy. Unfortunately, this kind of deep deliberation and policy construction can require significant computation time and is not suited for online tasks.
To aid in policy computation at runtime, this work proposes a two-phase approach to quickly reconfigure local control policies defined over small, discrete regions of the space in order to find a global control policy that is both robust to uncertainties and allows the stochastic system to achieve its task. Using a Bounded-parameter Markov Decision Process (BMDP), local control policies are computed within discrete regions offline, and one policy from each region is selected by the BMDP to achieve a sequence of tasks specified at runtime. Local policy selection is optimal with respect to a continuously-valued reward function, and experiments show that highly desirable control policies can be computed in seconds.