Learning Object Classifiers with Limited Human Supervision on a Physical Robot

In recent years, deep learning approaches have been leveraged to achieve impressive results in object recognition. However, such techniques are problematic in real world robotics applications because of the burden of collecting and labeling training images. We present a framework by which we can direct a robot to acquire domain-relevant data with little human effort. This framework is situated in a lifelong learning paradigm by which the robot can be more intelligent about how it collects and stores data over time. By iteratively training only on image views that increase classifier performance, our approach is able to collect representative views of objects with fewer data requirements for longterm storage of datasets. We show that our approach for acquiring domain-relevant data leads to a significant improvement in classification performance on in-domain objects compared to using available pre-constructed datasets. Additionally, our iterative view sampling method is able to find a good balance between classifier performance and data storage constraints.