Unsupervised Transfer Learning For Video Prediction Based on Generative Adversarial Network
Jiwen Shi, Qiuguo Zhu, Jun Wu · 2021 27th International Conference on Mechatronics and Machine Vision in Practice (M2VIP) · 2021
Video prediction technology offers new benefits for robots in completing a series of tasks based on self-supervised planning. In a video prediction task in a scene from the physical world, we must always attend to the authenticity of the prediction. However, the prediction model can model only the specific scene; when the scene changes, we need to train a new model from scratch to fit the domain distribution of the data, which often requires a large amount of specific data and long training time. Therefore, ways to minimize the constraints on model learning for new senses is an important issue. Aiming at its authenticity and generalization, this paper examines video prediction based on adversarial training and unsupervised transfer learning. For the interactive operation of a manipulator, a method of video prediction based on a generative adversarial structure using the transfer learning strategy is proposed. Using the Sawyer manipulator strong dataset as a guide and a UR5 manipulator dataset as target, the results of experiments verified that the prediction generator trained by the adversarial method can generate realistic results, and the transfer learning strategy has the following two advantages: (1) The number of training iterations for a new target scene is reduced, convergence is faster, overall training time is shorter, and thus efficiency is higher. (2) The prediction of interactive objects has better generalization capability.