Experiments on learning-based industrial bin-picking with iterative visual recognition

This paper aims to automate the picking task needed in robotic assembly. Parts supplied to an assembly process are usually randomly staked in a box. If randomized bin-picking is introduced to a production process, we do not need any part-feeding machines or human workers to once arrange the objects to be picked by a robot. The authors introduce a learning-based method for randomized bin-picking.

The authors combine the learning-based approach on randomized bin-picking (Harada et al., 2014b) with iterative visual recognition (Harada et al., 2016a) and show additional experimental results. For learning, we use random forest explicitly considering the contact between a finger and a neighboring object. The iterative visual recognition method iteratively captures point cloud to obtain more complete point cloud of piled object by using 3D depth sensor attached at the wrist.

Compared with the authors’ previous research (Harada et al., 2014b) (Harada et al., 2016a), their new finding is as follows: by using random forest, the number of training data becomes extremely small. By adding penalty to occluded area, the learning-based method predicts the success after point cloud with less occluded area. We analyze the calculation time of the iterative visual recognition. We furthermore make clear the cases where a finger contacts neighboring objects.

The originality exists in the part where the authors combined the learning-based approach with the iterative visual recognition and supplied additional experimental results. After obtaining the complete point cloud of the piled object, prediction becomes effective.