Learning to Grasp through World Models with Backward Prediction

In the era of the fourth industrial revolution, the paradigm of intelligent manufacturing has ascended as a pivotal driver for future advancements. The combination of intelligent manufacturing with reinforcement learning offers significant benefits and opportunities. However, applying reinforcement learning in practical settings faces challenges due to sampling efficiency. Recently, model-based reinforcement learning algorithms have demonstrated higher sample efficiency. In this research, we introduce an enhanced version of the Dreamer algorithm and devise a system dedicated to mastering the tasks of picking and placing through the limited data collected. The improved algorithm addresses the issue of low sample efficiency and proves to be exceptionally effective in pixel-based tasks associated with high-dimensional state spaces. To corroborate the efficacy of the proposed algorithm, we evaluate the algorithm on a simulation environment of a production line for robotic arm operations developed in Cop-peliasim. The merits of the augmented algorithm are further corroborated through its performance in complex DeepMind control tasks.