Abstracting from Robot Sensor Data using Hidden Markov Models

摘要

This work is the rst step of a larger e ort aimed at learning logical descriptions from robot sensory data. These representations are more compact than sensory traces and will support logical reasoning. We view the robot's experiences as trajectories through an unknown state space. The robot receives information about the state of the world through its sensors. We present a technique to automatically extract atomic propositions from these sensors. Our assumption is that a state means that something is invariant in the world, and that this invariance is re ected in some constant sensor values, or constant functions of sensor values. Our task is then to nd the states and their invariant characterization. We employ a hidden Markov model to nd the states and their distributional characterization. From the probability distributions of the sensor values in states we then create atomic propositions that describe the states. The transformation of atomic propositions into predicates should be straightforward given sensor models that specify the arguments of these predicates, but this has yet to be implemented.