Pose estimation of robotic manipulators using deep transfer learning towards video-based system identification

Abstract

Pose estimation from high-dimensional inputs, such as video input, is a challenging system identification problem. In robotics an accurate model can provide better performance in model-predictive control (MPC). In this work, the identification input is the video of an original robotic manipulator link with a flexible joint and the angular position of the link: the output. The methodology is implemented using deep convolutional neural networks (ConvNet). In order to optimize the training, a Transfer Learning (TL) approach is implemented, using five pre-trained ConvNets classification models. The estimation made by the deep ConvNets and the measured angular position from the original assembly. The TL method shows an important ability to estimate the angular position of the link from the frames of the video, with every model obtaining over 0.98 determinant coefficient score. The VGG19 model shows the best accuracy at estimating the link’s position from the frames of the video, decreasing the Mean Absolute Error (MAE) by 25.71% compared to the second-best model.