[1G5-OS-21b] 世界モデルと知能

[1G5-OS-21b-02] Goal-conditioned self-supervised learning from play for flexible object manipulation

〇Keigo Ishii¹, Shun Hiramatsu¹, Yuta Nomura¹, Shingo Murata¹ (1. Keio Univ.)

Keywords:Robot learning, Deep learning, Self-supervised learning, Flexible object manipulation, Play data

The demand for robots capable of manipulating flexible objects such as cables and fabrics is increasing in production and daily life environments. However, modeling the state of these objects is challenging due to their infinite shapes. Therefore, a data-driven deep learning approach has been identified as an effective solution. While reinforcement learning and imitation learning are two specific methods, each has its limitations. Reinforcement learning requires extensive data for exploration, making it inefficient in terms of data collection. On the other hand, imitation learning usually requires labeling (task specification), which restricts the robot's possible action patterns. To address these limitations, this study proposes a new framework for robotic flexible object manipulation. The framework collects play data by operating a robot with human curiosity and performs goal-conditioned self-supervised learning using extracted sub-sequences from the play data. The effectiveness of the proposed framework is demonstrated through robot experiments on rope manipulation.