Feature semantic space-based sim2real decision model

Wenwen Xiao, Xiangfeng Luo, Shaorong Xie

Abstract: At present, the intelligent decision model of unmanned systems can only be applied to virtual scenes, which makes it difficult to migrate to real scenes because the image gap between virtual scenes and real scenes is relatively large. The main solutions are domain randomization, domain adaptation, and image translation. However, these methods simply add noise and transform the perceptual information and do not consider the semantic information of the agent’s perceptual space. This causes the problem of low accuracy in the migration of virtual scene decision models to real scenes. Considering the above problems, we propose a feature semantic space-based sim2real decision model, which includes an environment representation module, policy optimization module and intelligent decision module. The model framework can narrow the image gap between real-world scenes and virtual scenes. First, using the environment representation module, the virtual scene and real scene are simultaneously mapped to the feature semantic space through semantic segmentation. Then, in the policy optimization module, we propose an AMDDPG policy optimization algorithm. The algorithm obtains the local and global experience in the learning process through the global and local network architecture. It also solves the problem of the slow learning rate of sim2real. Finally, in the intelligent decision module, the data in the semantic space integrating virtual scene and real scene features are used as the training data of the agent autonomous decision model. Experimental results confirm that our method has more effective generalization and robustness of the model in the real scene and can be better migrated to the real scene.