Master student in robotics, computer science or computer graphics (f/m/x): Inter-robot 6D articulated pose estimation for multi-robot outdoor SLAM
What to expect:
Simultaneous Localization and Mapping (SLAM) is a critical skill for a mobile robotic system to operate autonomously in a GNSS-denied environment, such as planetary bodies. When a team of robots, as opposed to a single agent, is performing collaborative SLAM, a necessary ability is the one to detect the other systems belonging to the team using camera images, and estimate a relative pose between them. This translates directly into the SLAM graph as a pose constraint that allows to join the maps from the respective systems into a combined representation. Traditionally, 6D pose estimation is performed by observing markers (e.g. AprilTag) rigidly attached to each one’s body. This approach severely limits the range at which systems can detect each other, thus introducing the need of “marker-less” solutions.
The objective of this thesis work is to design a “marker-less” approach, based on prior knowledge at the Institute of Robotics and Mechatronics, where the pose of robots is detected in each other’s image as an ensemble of articulated objects. Plausibility metrics, based on a-priori knowledge about the robot’s joint states, should be used if available. Focus will be put on the development of a working ROS2 node, that should be deployed on the real robotic platform, as well as ad-hoc simulation frameworks, based on existing tools developed within the department, for the training of the network.
The developed approach should be both tested on data recordings from analogous mission campaigns, as well as deployed on the real robotic system and tested in “real-time”. Towards the collection of data for offline testing, it is foreseen the usage of the DLR Outdoor planetary exploration laboratory, equipped with an outdoor Vicon system for the collection of ground truth poses.
What we expect from you:
- Bachelor’s degree in Computer Science, Electrical Engineering, Robotics Engineering or a similar
- Knowledge on Deep Learning approaches for image processing, and experience with the PyTorch library
- Solid knowledge on image processing techniques and pattern recognition (e.g. feature detection and matching)
- Very good Python, and good C++ coding skills
- Fluency with the English language
- Intent to disseminate the results of this work through publications in relevant computer vision or robotics conferences
To apply for this position, please include a resume and transcript of grades.
Further information:
Starting date: Immediate
Duration of contract: 6 months