Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Goal-driven Exploration of Unknown Spaces using Landmarks and Semantic Relationships
The goal of this project is to develop a vision-based learning method that allows an autonomous agent to explore a previously unknown space with the goal of discovering a queried semantic. This task can become particularly complex in unknown large-scale environments. A way to decrease the complexity, path traveled, and time needed to complete the task, is to provide the agent with a prior understanding about how the world is structured. For example, when asked to discover all the chairs in an apartment, the agent can benefit from prior knowledge of where it is most probable to find the queried semantic (e.g., next to a desk or a dining table). Our premise is that, if the agent can identify while exploring the space a larger and more prominent landmark that is commonly found close to the queried semantic, it will complete the task faster and having traveled a shorter path. In this project, we will leverage object co-occurences and other semantic relationships during learning to perform goal-driven exploration in simulation using existing indoor real-world datasets. As part of the project, we will test generalization of the method in several unseen environments, hence it is important to devise a method that does not rely on the metric space.
Not specified
Not specified
Daniel Barath (dbarath@ethz.ch), Iro Armeni (armeni@ibi.baug.ethz.ch)
Daniel Barath (dbarath@ethz.ch), Iro Armeni (armeni@ibi.baug.ethz.ch)