Optimizing Relocalization Likelihood in Challenging Environments With Reinforcement Learning

Sammanfattning

If an autonomous agent lost track of its global pose while moving through an environment, it would require the ability to perform global localization, called relocalization,autonomously to find its global pose. Standard autonomous methods struggled with relocalizing in difficult factory and mine environments due to their symmetry and lack of unique features, leading to the need for another process that could relocalize within such environments. An approach was to use reinforcement learning since it could learn how and where the agent should move to optimize the relocalization likelihood. The relocalization problem was made into a reinforcement learning problem by reformulating it as a Markov decision problem based on the Monte Carlo localization algorithm. Relocalization had been achieved once the Monte Carlo localization’s particles converged, their spread infuenced by how the agent moved and what its sensors measured from where it had moved. This meant a proximal policy optimization algorithm could be trained on it to create the agent’s movement policy. The reinforcement learning algorithm showed great potential and, on some maps, managed to even outperform the explore-exploit baseline created for this task.