Master Thesis - Out-of-distribution detection + annotation for traversability estimation for robots

Advertisement for the field of study such as: Automation technology, electrical engineering, computer science, cybernetics, mechanical engineering, mathematics, mechatronics, control engineering, software design, software engineering, technical computer science or comparable.

In the Professional Service Robots - Outdoor research group we develop autonomous, mobile robots for a variety of outdoor applications, such as agriculture, forestry and logistics. The focus is on the development of an autonomous outdoor navigation solution as well as the hardware of the robots.

For mobile robots operating in outdoor, unstructured environments with unknown terrain conditions, accurately understanding the traversability of the surrounding environment is essential. This ensures that the robot navigates through safe paths, avoiding difficult terrains such as mud and dense vegetation and preventing collision with obstacles.

A key approach in this research field consists of using Deep Neural Networks and Foundation Models to perform traversability inference on incoming camera images. However, a major challenge lies in the lack of extensive datasets for the different field environments and the high effort in acquiring and labeling the data set. This is especially limiting for robots operating in a wide variety of environments and performing exploration tasks.

Therefore, state-of-the-art semantic traversability classification algorithms need to be extended with the capability to detect out-of-distribution environments and to autonomously infer their traversability.

In this thesis, you will design a semantic traversability classification DNN pipeline capable of performing long-term traversability estimation. You will focus on extending our state-of-the-art few-shot segmentation DNN algorithm to enable automatic adaptation to new environments. In particular, you will evaluate different methods to identify out-of-distribution domains. You will also evaluate different approaches for automated model-learning using the robot’s experience in these new environments. You will test your implementation in real-world scenarios using both recorded data and real-life deployment in our mobile CURT robots.