Automated Benchmarks and Optimization of Perception Tasks

Abstract

Advanced robots operating in complex and dynamic environments require intelligent perception algorithms to navigate collision-free, analyze scenes, recognize relevant objects, and manipulate them. Nowadays, the perception of mobile manipulation systems often fails if the context changes due to a variation e.g. in the lightning conditions, the utilized objects, the manipulation area, or the environment. Then, a robotic expert is needed who needs to adjust the parameters of the perception algorithm and the utilized sensor or even select a better method or sensor. Thus, a high-level cognitive ability that is required for operating alongside humans is to continuously improving their performance based on introspection. This adaptability to changing situations requires different aspects of machine learning, e.g. storing experiences for life-long learning, creating annotated datasets for supervised learning through user interaction, Bayesian optimization to avoid brute-force search in high-dimensional data, and a unified representation of data and meta-data to facilitate knowledge transfer. Here we present how we automated and integrated different aspects of these.