The learning technologies for picki

The learning technologies for picking tasks presentedin the preceding sections use images and convolutionalneural networks. Robot sensors such as force sensors andtactile sensors mostly provide time-series information.Among the networks effective for learning from suchtime-series information, the Recurrent Neural Network(RNN) [45] and Long Short-Term Memory (LSTM) [46]have been widely used to make industrial robots more in-telligent. Hatori et al. connected an LSTM-based voicerecognition network to their robot and demonstrated bin-picking based on voice interactions [47]. Yang et al.demonstrated their robot’s RNN-based learning to fold uptowels in many different colors and shapes [48]. Gu et al.had their robot learn to open and close a door based on itsreinforcement learning [49]. Finn et al. had their robot ac-quire motions to shift picked objects sideways [50]. Ref-erence [51] succeeded in a robot learning a task in aninvisible state hidden from its vision sensor, combinedwith a tactile sensor. Machine learning is sure to becomean increasingly important technology for robotic object-manipulation tasks.