Wrong today, right tomorrow: Experience-based classification for robot perception

Abstract

This paper is about building robots that get better through use in their particular environment, improving their perceptual abilities. We approach this from a life long learning perspective: we want the robot’s ability to detect objects in its specific operating environment to evolve and improve over time. Our idea, which we call Experience-Based Classification (EBC), builds on the well established practice of performing hard negative mining to train object detectors. Rather than cease mining for data once a detector is trained, EBC continuously seeks to learn from mistakes made while processing data observed during the robot’s operation. This process is entirely self-supervised, facilitated by spatial heuristics and the fact that we have additional scene data at our disposal in mobile robotics. In the context of autonomous driving we demonstrate considerable object detector improvement over time using 40Km of data gathered from different driving routes at different times of year.