Bridging the Last Mile in Sim-to-Real Robot Perception via Bayesian Active Learning

Learning from synthetic data is popular in avariety of robotic vision tasks such as object detection, becauselarge amount of data can be generated without annotationsby humans. However, when relying only on synthetic data,we encounter the well-known problem of the simulation-to-reality (Sim-to-Real) gap, which is hard to resolve completelyin practice. For such cases, real human-annotated data isnecessary to bridge this gap, and in our work we focus on howto acquire this data efficiently. Therefore, we propose a Sim-to-Real pipeline that relies on deep Bayesian active learningand aims to minimize the manual annotation efforts. We devisea learning paradigm that autonomously selects the data thatis considered useful for the human expert to annotate. Toachieve this, a Bayesian Neural Network (BNN) object detectorproviding reliable uncertain estimates is adapted to infer theinformativeness of the unlabeled data, in order to performactive learning. In our experiments on two object detectiondata sets, we show that the labeling effort required to bridge thereality gap can be reduced to a small amount. Furthermore, wedemonstrate the practical effectiveness of this idea in a graspingtask on an assistive robot.