Flexible Object Manipulation by a Dual-Arm Robot Using Deep Predictive Learning: Near-Future Prediction and Real-Time Motion Generation

Abstract

In this study, we propose flexible object manipulation by a humanoid robot using deep predictive learning, which can respond to unlearned environments and work objects by predicting real-time actions suitable for the real world based on past learning experience. For the towel-hanging task as a flexible object manipulation, the robot grasps a towel placed on a desk and hangs it on a clothesline. The robot predicts the near-future situation based on visuomotor information, generates actions to minimize the error from reality, and continues to adjust its actions in real time while tolerating the difference between learning and reality, enabling the robot to work flexibly even in unlearned situations.