In this research project, we investigate approaches for autonomous semantic perception in partially observable settings, where the state of the agent as well as the (geometric and semantic) state of the environment are unknown. Specifically, our approach models the inherent coupling between visual scene appearance and the viewpoint. This coupling is represented by a viewpoint-dependent classifier model. Further, due to this coupling, the continuous and discrete variables are not statistically independent, and therefore, a joint probability distribution over these variables has to be maintained. The project focuses on probabilistic inference, planning under uncertainty, and learning in these settings.
Prerequisites:
- Strong programming skills (preferably Python or C++). Background in (deep) reinforcement learning, computer vision, robotics is an advantage.
Academic supervisor:
- Prof. Vadim Indelman (email)
Duration: 1 or 2 semesters