Embodied Active Domain Adaptation for Semantic Segmentation via Informative Path Planning

This work presents an embodied agent that can adapt its semantic segmentation network to new indoor environments in a fully autonomous way. Because networks fail generalize well unseen environments, the collects images of environment which are then used for self-supervised domain adaptation. We formulate this as informative path planning problem, and present novel information gain leverages uncertainty extracted from model safely collect relevant data. As adaptation progresses, these uncertainties change over time rapid learning feedback our system drives different Experiments show method adapts faster with higher final performance compared exploration objective, successfully be deployed real-world on physical robots.