1. Learning Re-grabbing Policies based on Grabbed Garbage Weight Estimation using In-bucket Images for Waste Cranes
- Author
-
Sasaki, Hikaru, Watanabe, Go, Hirabayashi, Terushi, Kawabata, Kaoru, and Matsubara, Takamitsu
- Abstract
The automation of waste cranes has been demanded to perform garbage incineration work with fewer workers efficiently. In particular, data-driven learning approaches are desirable for waste crane automation. When deciding whether to re-grab garbage by a waste crane to grab more garbage, human operators are efficiently making decisions based on visual information. However, the current automation system has to lift up the bucket to measure grabbed garbage weight to decide re-grab. The lifting motion after grabbing makes the crane motion inefficient. For this limitation, we propose a re-grabbing decision system with feedback from in-bucket camera images for the efficiency of waste crane automation by introducing the vision sensor in the bucket. To simplify the decision process of re-grabbing, we separate the re-grabbing decision system from the in-bucket image into grabbed garbage weight estimation and the re-grabbing decision policy based on estimated garbage weight. Moreover, the weight estimator model and the re-grabbing policy model are designed based on a Bayesian manner for data efficiency. The effectiveness is verified in a robotic waste crane system with an in-bucket camera. We confirmed the proposed method could learn re-grabbing decisions from the autonomously collected data. It achieved more efficient re-grabbing by waste cranes than the conventional automated system.
- Published
- 2023
- Full Text
- View/download PDF