Self-localization during navigation with noisy sensors in an ambiguous world is computationally challenging, yet animals and humans excel at it. In robotics, {\em Simultaneous Location and Mapping} (SLAM) algorithms solve this problem through joint sequential probabilistic inference of their own coordinates and those of external spatial landmarks. We generate the first neural solution to the SLAM problem by training recurrent LSTM networks to perform a set of hard 2D navigation tasks that require generalization to completely novel trajectories and environments. Our goal is to make sense of how the diverse phenomenology in the brain's spatial navigation circuits is related to their function. We show that the hidden unit representations exhibit several key properties of hippocampal place cells, including stable tuning curves that remap between environments. Our result is also a proof of concept for end-to-end-learning of a SLAM algorithm using recurrent networks, and a demonstration of why this approach may have some advantages for robotic SLAM.

An LSTM recurrent neural network is trained to perform simultaneous localization and mapping (SLAM) tasks, given noisy odometry data and occasional input from contacts with walls. The paper is well written and I find it interesting to study recurrent neural network solutions of SLAM. But the idea of using LSTM-RNNs for SLAM problems is not new (as the authors mention in the discussion, line 256) and the results do not really surprise me. Given the amount of training trials, the network has presumably seen sufficiently many trajectories (and variations of the environment) to perform well on most test examples (and if the difference between training and testing error is too large, follow the standard recipes and tune the number of hidden neurons and the number of training trials). A more interesting result would be it the LSTM-RNN would generalize to out of distribution samples, i.e. if its performance would be comparable to that of a PF SLAM in totally different environments than the ones seen during training.

Self-localization during navigation with noisy sensors in an ambiguous world is computationally challenging, yet animals and humans excel at it. In robotics, {\em Simultaneous Location and Mapping} (SLAM) algorithms solve this problem through joint sequential probabilistic inference of their own coordinates and those of external spatial landmarks. We generate the first neural solution to the SLAM problem by training recurrent LSTM networks to perform a set of hard 2D navigation tasks that require generalization to completely novel trajectories and environments. Our goal is to make sense of how the diverse phenomenology in the brain's spatial navigation circuits is related to their function. We show that the hidden unit representations exhibit several key properties of hippocampal place cells, including stable tuning curves that remap between environments.