[2G6-OS-21f] 世界モデルと知能

[2G6-OS-21f-02] Real-World Robot Control and Data Augmentation by World Model Learning from Play

〇Yuta Nomura¹, Shingo Murata¹ (1. Keio Univ.)

Keywords:Play Data, World Model, Learning from Demonstration, Robot Learning

Robots are expected to perform various tasks in complex environments with a high generalization ability, similar to that of humans. Generally, imitation learning with expert demonstrations has high efficiency but low generalization ability. In contrast, reinforcement learning with explorations has high generalization ability but low efficiency. To combine their strengths, we focus on ``play data'' collected by humans teleoperating a robot with curiosity. Specifically, we propose a framework for real-world robot control and data augmentation based on world model learning from play data. Robot experiments demonstrated that the robot with the framework can perform goal-conditioned object manipulation tasks. Furthermore, we also found that simulation in the world model can create novel combinations that are not included in the original play data. These findings suggest that further learning the augmented data has the potential to enable the robot to acquire higher generalization ability.