Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Decentralized Place Recognition for Multi-Robot SLAM
This project targets the exploration multi-agent collaborative Simultaneous Localization And Mapping (SLAM) using a decentralized paradigm. We are particularly interested in prototyping a scalable decentralized place recognition module in order to establish inter-agent loop closures.
Keywords: Collaborative SLAM, Decentralized SLAM, Place Recognition, Loop Closure, Bag-of-Words
SLAM algorithms designed for use on a single robot have seen tremendous progress in terms of accuracy and computational performance over the past decade. With the relative maturity of single agent SLAM, recent research directions address the challenges of fusing information from a team of robots in an effort to enhance the robustness and accuracy of the estimate of each robot in the team. Approaches following a centralized collaboration paradigm, for instance, aim to offload expensive computations to a central server entity that coordinates and augments the estimate of each robot in the team. However, in situations where reliance on a central server cannot be guaranteed (e.g. search and rescue missions) or in the case of large team size (e.g. drone formation flying), decentralized approaches are advantageous.
Given the promise of the decentralized paradigm, in this project we aim to investigate solutions to decentralized place recognition, which is an essential module used to establish correspondences between agents. In particular, the focus of this project is on approaches that are scalable with respect to communication bandwidth requirements while retaining sufficient recall.
Students interested in this topic are encouraged to contact the supervisors to arrange an individual discussion about a potential thesis.
SLAM algorithms designed for use on a single robot have seen tremendous progress in terms of accuracy and computational performance over the past decade. With the relative maturity of single agent SLAM, recent research directions address the challenges of fusing information from a team of robots in an effort to enhance the robustness and accuracy of the estimate of each robot in the team. Approaches following a centralized collaboration paradigm, for instance, aim to offload expensive computations to a central server entity that coordinates and augments the estimate of each robot in the team. However, in situations where reliance on a central server cannot be guaranteed (e.g. search and rescue missions) or in the case of large team size (e.g. drone formation flying), decentralized approaches are advantageous. Given the promise of the decentralized paradigm, in this project we aim to investigate solutions to decentralized place recognition, which is an essential module used to establish correspondences between agents. In particular, the focus of this project is on approaches that are scalable with respect to communication bandwidth requirements while retaining sufficient recall. Students interested in this topic are encouraged to contact the supervisors to arrange an individual discussion about a potential thesis.
- Literature review and familiarization with (decentralized) place recognition.
- Identification of a promising approach and proof-of- concept implementation in MATLAB or Python.
- Evaluation and comparison to state-of-the art methods.
- (Optional) Integration of the approach into a larger
framework in C++.
- Literature review and familiarization with (decentralized) place recognition. - Identification of a promising approach and proof-of- concept implementation in MATLAB or Python. - Evaluation and comparison to state-of-the art methods. - (Optional) Integration of the approach into a larger framework in C++.
- Strong analytical skills.
- Experience in some or all of the following programming languages: MATLAB, Python, C/C++.
- Familiarity with some or all of the following is desirable: Linux, Eigen, OpenCV, Ceres and/or similar frameworks.
- Prior knowledge about SLAM and Place Recognition is advantageous.
- Strong analytical skills. - Experience in some or all of the following programming languages: MATLAB, Python, C/C++. - Familiarity with some or all of the following is desirable: Linux, Eigen, OpenCV, Ceres and/or similar frameworks. - Prior knowledge about SLAM and Place Recognition is advantageous.
Philipp Bänninger (baephili@ethz.ch)
Marco Karrer (marco.karrer@mavt.ethz.ch)
Philipp Bänninger (baephili@ethz.ch) Marco Karrer (marco.karrer@mavt.ethz.ch)