Navigation and planning in latent maps

We investigate the problem of agents planning paths to efficiently traverse environments only observed through the agent’s limited and local sensors such as laser range finders. We use a probabilistic sequential latent-variable model with an explicit global map. The entire model is trained only from local sensor readings, leveraging recent advances in approximate Bayesian inference. The generative nature of the model allows us to plan traversals of the real environment solely based on simulations within the learnt model. Further, we can exploit the Euclidean geometry imposed on our map to make use of well-known path planning algorithms. We showcase this ability by reliably and efficiently finding optimal paths in random mazes.