Computer Vision News - December 2021

6 Robot Learning Research The first element is based on previous work called InfoBot: Transfer and Exploration via the Information Bottleneck which postulates an information bottleneck architecture. The graphical model on the right explicates the architecture: the encoder extracts a representation of the goal image , conditioned on the current observation . The encoded representation, which is built such as to only retain relative location of the goal from the context, is then decoded into a prediction of the best action and the temporal distance to the goal. The objective of the model consists into maximising the model’s ability to predict actions and distances from the encoded representation, and the model’s compression of the incoming goal image. This architecture is used to pre-train a latent goal model on the offline dataset described before, which is used as building block for the second component of RECON: the topological map. This is created during the exploration phase and then employed again in the navigation one, to quickly navigate towards the goal . The map’s topological memory is built incrementally with exploration by looking for subgoals, which are represented by the latent variables in the model. Given a subgoal, which is chosen based on the robot’s estimate of the goal reachability and its proximity to the frontier, the algorithm executes actions and its distance to the goal is used to construct edges of the topological graph. More implementation details (hyperparameters and architecture) and algorithmic components which are at the core of the RECON method (exploration techniques) are also described by the authors in great details as part of the supplementary and provided as pseudo codes.