@ai_and_law · Post #191 · 17.12.2023 г., 09:43
🌟 AI Sunday Wonders: Teaching Robots Household Tasks in Minutes Hello everybody! Welcometo AI Sunday Wonders! Today, we dive into an exciting breakthrough that could revolutionize how robots learn to perform everyday tasks at home. Imagine a robot capable of mastering household chores in just 20 minutes. Well, meet Dobb-E, an open-source system that's shaking up the field of robotics by addressing a significant challenge—limited training data. Dobb-E was trained in real homes, learning tasks like opening doors and arranging cushions. Unlike AI models relying on internet data, Dobb-E's data was physically collected. Volunteers, armed with iPhones attached to grabber sticks, performed tasks in homes. The data collected, including movement and depth, was used to train Dobb-E. After just 13 hours of recordings, the team trained an AI model using self-supervised learning techniques. They tested it on a commercial robot called Stretch, which successfully executed tasks in real homes. Dobb-E achieved an 81% success rate in completing 109 household tasks. While Dobb-E showed promise, reflective surfaces posed challenges, and certain tasks were deemed risky. Despite this, the project is seen as tangible progress in home robotics. The NYU research team has made the project open source, inviting others to use the code and contribute to expanding the range of tasks Dobb-E can perform. The goal? A future where robots enter new homes already equipped with the knowledge of common tasks. #AISunday#RoboticsInnovation#DobbE#AIRevolution