Latent belief space planning using a trajectory tree

Abstract

Techniques for latent belief space planning include: during execution of an autonomous agent configured to control operation of a physical mechanism, obtaining a current observation of a physical environment; based at least on the current observation of the physical environment, generating a trajectory tree that represents possible trajectories in a belief space, wherein nodes of the trajectory tree represent values of a continuous observation, a continuous state, and a continuous control, each node being associated with one of multiple timesteps along the plurality of possible trajectories, and wherein branches from inner nodes to child nodes correspond to possible outcomes and observations of a multi-modal latent state; determining a current value of the continuous control associated with a current node; and applying the current value of the continuous control to operation of the physical mechanism.