Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Deep Learned Graph Clustering
In this project we develop a method to clusster novel object categories without any supervision from a robot's observations of its environnment. Unknown objects that a robot finds should be grouped by their similarities. We aim for an end-to-end differentiable solution that leverages newest research from representation learning and graph processing.
The overarching goal of this project is to enable robots that can continually and autonomously learn from their environment by adapting to novel environments and learn to identify new object categories. The established approach of supervised learning on large datasets always has problems with domain and sim-to-real gaps. This line of work therefore represents a shift towards learning during the robot’s mission, naturally removing any domain gaps but also reducing the amount of supervision that can be applied.
In this particular project, we look at the problem of clustering observations to automatically discover object categories. The goal of this work is to develop clustering-based learning methods that enable robots to explore a novel scene and learn from gathered data over a trajectory.
In the end, we want to have an end-to-end learnable system that can identify novel parts in a scene and use clustering to sort these novel parts into different classes. Building up on an existing proof -of-concept (https://arxiv.org/abs/2206.10670) and literature in hierarchical and self-supervised learning, we want to innovate deep learning methods for a closer integration of clustering and learning.
The overarching goal of this project is to enable robots that can continually and autonomously learn from their environment by adapting to novel environments and learn to identify new object categories. The established approach of supervised learning on large datasets always has problems with domain and sim-to-real gaps. This line of work therefore represents a shift towards learning during the robot’s mission, naturally removing any domain gaps but also reducing the amount of supervision that can be applied. In this particular project, we look at the problem of clustering observations to automatically discover object categories. The goal of this work is to develop clustering-based learning methods that enable robots to explore a novel scene and learn from gathered data over a trajectory. In the end, we want to have an end-to-end learnable system that can identify novel parts in a scene and use clustering to sort these novel parts into different classes. Building up on an existing proof -of-concept (https://arxiv.org/abs/2206.10670) and literature in hierarchical and self-supervised learning, we want to innovate deep learning methods for a closer integration of clustering and learning.
- Familiarize yourself with the literature
- Set up a minimal experiment
- evaluate different basic ideas
- Iterate by increasing complexity and bringing in your own ideas
- (bonus) integrate into existing system for on-robot experiments
- Familiarize yourself with the literature - Set up a minimal experiment - evaluate different basic ideas - Iterate by increasing complexity and bringing in your own ideas - (bonus) integrate into existing system for on-robot experiments
- Highly dedicated and motivated student
- Familiar with python and pytorch
- prior practical experience with deep learning is a plus
- Highly dedicated and motivated student - Familiar with python and pytorch - prior practical experience with deep learning is a plus
Please send your CV and transcript to blumh@ethz.ch.
Please send your CV and transcript to blumh@ethz.ch.