Probabilistic Robot Action Planning/Control

Robot behaviors are often designed using a set of states, with transition conditions among them, i.e. a finite state machine. The transition conditions often depend on uncertain inputs. Thus, transitions could fire incorrectly or fail to fire. This necessitates complicating the behavior design with additional transitions to “back out” of states judged to be erroneous. Current robotic systems maintain probabilistic models of their environment. Deterministic state transitions are a mismatch for these models in that they require a single datum to be extracted from the probabilistic models to use as an input to the state transition decision. For example, a particle filter would be processed to produce a choice for “the” current robot pose to use as a basis for action decisions. In some sense, this “throws away” the advantages of tracking multiple hypotheses in the particle filter. In this work, we seek to carry the probabilistic models into behavior decision making, insofar as it is possible. Clearly, a robot cannot choose to, for example, move to two different places at once, but not all decisions are mutually exclusive. For example, a robot could move its head to visually scan for objects while walking to a target location.