Repeatable Folding Task by Humanoid Robot Worker Using Deep Learning

Learning deep movement primitives using convolutional neural networks

“…Our approach requires relatively fewer training data for learning, as compared with other CNN based motor learning approaches [7], [8], [18]. Applying RL for learning a motor skill can require a lot of trials [8].…”

“…Since we want to use camera images, if a human is always present in the images during demonstrations, a CNN can learn human specific feature. Now during motion reproduction, if the human is not present in the image, then it can result in a failure of the task during reproduction phase [7], [8], [18]. Alternatively a human can provide teleoperated demonstrations as in [11], [18].…”

“…Now during motion reproduction, if the human is not present in the image, then it can result in a failure of the task during reproduction phase [7], [8], [18]. Alternatively a human can provide teleoperated demonstrations as in [11], [18]. Since generating teleoperated demonstrations can be a time consuming and a tedious task for a human operator, we use an already trained TP-DMP model for collecting the demonstrations.…”

“…We select 45 motions (21600 data-points) for the training set while the remaining 5 motions (2400 data-points) are used for validating the learned model. This is lower than 800,000 grasp attempts in [7], at least 156 execution trails in [8] and 24, 500 and 3500 data-points in training and validation sets respectively in [18]. RGB images were captured with a Kinect Xbox 360 camera.…”

“…These constraints limit the applicability of their approach to complex tasks. A PbD approach based on a combination of convolutional auto-encoders and a fully connected neural network is presented in [18]. A shortcoming of their approach is that it cannot achieve the generalization properties which are typically associated with a DMP, for instance temporal rescaling of the motion and the variation of the goal position.…”