Dobb·E is an open-source framework for teaching home robots to manipulate everyday objects using imitation learning in about 20 minutes. It targets real apartments and household routines rather than controlled lab setups, and reports a strong average task success rate.
How it works
Dobb·E collects human demonstrations with a simple physical tool called Stick. These demonstrations are used to build the Homes of New York (HoNY) dataset, designed to capture the variety of real home environments. The project then trains Home Pretrained Representations (HPR) to help robots form better environment-aware representations.
Fast training for new tasks
After pretraining, Dobb·E needs roughly five minutes of new data in an unfamiliar home for a new task. About fifteen minutes later, it produces a robot control policy that achieves around 81% average success.
Who it’s for
- Robotics researchers working on real-world manipulation
- Developers using imitation learning and reinforcement learning
- Anyone studying AI applications in household robotics
