Dobb-E is an innovative open-source framework designed to enhance the capabilities of household robots through imitation learning. This system addresses the common limitations in home robotics by utilizing an affordable and user-friendly tool called the Stick. Constructed from a $25 Reacher-grabber stick, 3D printed components, and an iPhone, the Stick facilitates the collection of valuable demonstration data. The framework leverages a comprehensive dataset known as Homes of New York (HoNY), which contains 13 hours of interactions from 22 different homes, enriched with RGB and depth videos and detailed action annotations.

The core of Dobb-E's functionality lies in its ability to train a representation learning model called Home Pretrained Representations (HPR), based on the ResNet-34 architecture. This model employs self-supervised learning to equip robots with the skills necessary to perform new tasks in unfamiliar environments. With an impressive average success rate of 81% in solving novel tasks within 15 minutes, Dobb-E provides access to pre-trained models, code, and comprehensive documentation via GitHub, along with a research paper detailing its methodology.