Multitask Learning for Scalable and Dense Multilayer Bayesian Map Inference

This article presents a novel and flexible multitask multilayer Bayesian mapping framework with readily extendable attribute layers. The proposed goes beyond modern metric-semantic maps to provide even richer environmental information for robots in single formalism while exploiting intralayer interlayer correlations. It removes the need robot access process from many separate when performing complex task, advancing way interact their environments. To this end, we design deep neural network attention mechanisms as our front-end heterogeneous observations multiple map layers simultaneously. Our back-end runs scalable closed-form inference only logarithmic time complexity. We apply build dense robotic including occupancy traversability Traversability ground truth labels are automatically generated exteroceptive sensory data self-supervised manner. present extensive experimental results on publicly available datasets collected by 3D bipedal platform show reliable performance different Finally, also discuss how current can be extended incorporate more such friction, signal strength, temperature, physical quantity concentration using Gaussian software reproducing presented or running customized is made available.