MLNav: Learning to Safely Navigate on Martian Terrains

Towards a Quantitative Analysis of Martian Terrains

This document investigates the use of sample based Bayesian segmentation algorithms on images of the Martian landscape, with regard to a quantitative understanding of statistical and systematic errors. The errors predicted from stochastic sampling are estimated using the conventional methods of error propagation and the Cramer-Rao bound and compared to the estimation errors for computed total q...

Origin of the martian chaotic terrains.

Towards a Quantitative Segmentation of Martian Terrains

There is demand within the planetary science community for automated image analysis tools. Researchers in these areas commonly require estimated quantities of geological features appearing within satellite images for comparison against theoretical models. Such comparisons can only be meaningful if quantities are accompanied by appropriate estimates of uncertainty (e.g. error bars). Here, we sho...

Learning to Navigate in Complex Environments

Learning to navigate in complex environments with dynamic elements is an important milestone in developing AI agents. In this work we formulate the navigation question as a reinforcement learning problem and show that data efficiency and task performance can be dramatically improved by relying on additional auxiliary tasks. In particular we consider jointly learning the goal-driven reinforcemen...

Learning to navigate: experience versus maps.

People use "route knowledge" to navigate to targets along familiar routes and "survey knowledge" to determine (by pointing, for example) a target's metric location. We show that both root in separate memories of the same environment: participants navigating through their home city relied on representations and reference frames different from those they used when doing a matched survey task. Tüb...