Dobb-E
Teach household robots skills through imitation learning, a smart way to automate
Dobb-E: Revolutionizing Household Robotics
Dobb-E is an innovative, open-source framework that empowers the training of household robots through imitation learning. By addressing the limitations of current home robotics, Dobb-E provides an affordable and user-friendly solution for gathering demonstrations.
The Stick: A Game-Changer in Data Collection
To collect data from the Homes of New York (HoNY) dataset, Dobb-E utilizes The Stick - a clever tool crafted using a $25 Reacher-grabber stick, 3D printed parts, and an iPhone. This setup enables the gathering of RGB and depth videos, along with action annotations for the gripper's 6D pose and opening angle.
Home Pretrained Representations (HPR): A Key to Success
Dobb-E leverages this data to train a representation learning model known as Home Pretrained Representations (HPR), based on the ResNet-34 architecture and self-supervised learning objectives. This model initializes a robot policy for executing new tasks in unfamiliar environments.
Impressive Results: Dobb-E's Average Success Rate
Dobb-E has demonstrated an impressive 81% average success rate in solving novel tasks within 15 minutes, based on just five minutes of collected data in a new home.
Accessibility and Insights
For developers and researchers alike, Dobb-E offers access to pre-trained models, code, and documentation through GitHub. An open-access paper titled "On Bringing Robots Home" provides deeper insights into the methodology and results of Dobb-E.