Essay.06-PenglinCai-2100017825
Keywords: Goal representation; Vision; Language; Goal-Conditioned RL; Multi-Modality
Abstract: Representing goals has always been an important work in machine learning and agent systems. However, many end-to-end trained agents usually represents goals from a certain kind of perception, constructing the correspondence between sensory organs of humans and goal representations. This essay summarizes several commonly used methods of goal representation, from the perspective of vision, language, and auditory sense. In addition, some other previous work simply uses implicit embeddings or latent variables to represent goals, about which we will also discuss. As a combination of the aforementioned content, some recent work uses multi-modal representations of goals to convey much more information, which is worth considering for future researches.
Submission Number: 143
Loading