Distributed Active Learning for Semantic Segmentation on Walking Robots

Abstract

Quickly adapting to new and unknown environments is a vital functionality for autonomous robots. To increase their capabilities, they need a sophisticated self- and environmental awareness. Understanding the surroundings can be aided by a pixelwise semantic segmentation of the images taken by the robot. However, to achieve viable results a large database of annotated images is needed in advance. To aid in a quick understanding of the surroundings a distributed active learning approach for semantic segmentation is presented. The robot evaluates and selects images where it thinks annotation would lead to a better overall segmentation results. Since the robot is potentially on an autonomous remote mission, the communication might be limited. Therefore, the robot selects images which are stored in a buffer before they are transmitted in batches to the base station. The annotation of the images and training of a new model for the segmentation can be performed asynchronously to the robot's mission. The model of the robot is then updated once ready. Four different stream-based query metrics are deployed and tested on a dataset and on ANYmal in a real scenario. The proposed approach allows a robot to be deployed in an unknown surrounding and through quick and continuous learning it gains an understanding of the area.