Embodied Active Domain Adaptation for Semantic Segmentation via Informative Path Planning

TL;DR

This paper introduces an embodied agent that autonomously adapts its semantic segmentation to new indoor environments through informative path planning, leveraging uncertainty to efficiently collect data for self-supervised domain adaptation, demonstrated on real robots.

Contribution

It proposes a novel informative path planning approach using uncertainty for autonomous domain adaptation of semantic segmentation in indoor environments.

Findings

Faster adaptation to new environments compared to exploration objectives.

Higher final segmentation performance after adaptation.

Successful deployment on real-world robotic platforms.

Abstract

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