AuDaCITI

Recent achievements of AI were led by large computational models and equally large datasets, often requiring intensive human labor to curate and annotate. In some cases however, the AI can interact with its environment and users to autonomously generate its own learning data. In this project, we study a specific setting of a quadruped robot equipped with a gripper arm, allowing the robot to navigate a room and manipulate its objects. Through the use of reinforcement learning, the robot learns to proactively generate a representative dataset of the different object categories in the room, which it is then able to retrieve when queried. The project will be in collaboration with Aalto University, Finland, that will bring its expertise in robotics and computer vision, while Inria Scool will bring its expertise in reinforcement learning. By solving a traditionally supervised learning task in full autonomy, we hope to get insights on how AIs of the future could proactively seek to improve their modeling of the world.