Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Contrastive learning for navigation in the wild
Implementation of a contrastive learning navigation pipeline for ANYmal. Development of a novel algorithm for navigation based on expert demonstration and integration into the existing ANYmal software stack.
Deploying our legged robot ANYmal in natural outdoor environments requires estimation of high quality paths for safe and reliable navigation. In this project, we want to learn safe navigation from expert demonstrations. While geometrical planning approaches have shown to work well in structured indoor environments where the sensed geometry aligns with the support surface, these approaches fail in outdoors scenarios. In natural scenarios varying vegetation density leads to unreliable depth sensor readings and more complex semantics are needed to differentiate between terrains e.g. concrete and loose gravel. For drones purley geometrical planning is a good choice given that mostly all environment interactions will lead to a complete failure of the mission for legged robots the problem of traversability assessment is significantly more complex given the constant environment interactions.
Therefore we want to learn to navigate outdoors like humans but do not want to exactly mimic the paths chosen by a human given that for most scenarios multiple trajectories are good even though only a single trajectory is demonstrated. A promising approach to overcome this is applying contrastive learning and learning the semantic connection between image data and a chosen path by the human expert. In this Master`s Thesis we want to implement a full navigation pipeline based on contrastive learning.
Deploying our legged robot ANYmal in natural outdoor environments requires estimation of high quality paths for safe and reliable navigation. In this project, we want to learn safe navigation from expert demonstrations. While geometrical planning approaches have shown to work well in structured indoor environments where the sensed geometry aligns with the support surface, these approaches fail in outdoors scenarios. In natural scenarios varying vegetation density leads to unreliable depth sensor readings and more complex semantics are needed to differentiate between terrains e.g. concrete and loose gravel. For drones purley geometrical planning is a good choice given that mostly all environment interactions will lead to a complete failure of the mission for legged robots the problem of traversability assessment is significantly more complex given the constant environment interactions. Therefore we want to learn to navigate outdoors like humans but do not want to exactly mimic the paths chosen by a human given that for most scenarios multiple trajectories are good even though only a single trajectory is demonstrated. A promising approach to overcome this is applying contrastive learning and learning the semantic connection between image data and a chosen path by the human expert. In this Master`s Thesis we want to implement a full navigation pipeline based on contrastive learning.
- Literature Review on Contrastive Learning for Robotics
- Implementation of full network architecture and learning pipeline
- Implementation of navigation pipeline
- Analysis and evaluation of the pipeline
- Deployment on robot ANYmal and rigors field testing
- Literature Review on Contrastive Learning for Robotics - Implementation of full network architecture and learning pipeline - Implementation of navigation pipeline - Analysis and evaluation of the pipeline - Deployment on robot ANYmal and rigors field testing
- Strong coding skills (Required: Python)
- Deep learning basics (Preferred: PyTorch)
- Strong communication skills and motivation (Requirement)
- Prior experience in ROS (Optionally)
- Strong coding skills (Required: Python) - Deep learning basics (Preferred: PyTorch) - Strong communication skills and motivation (Requirement) - Prior experience in ROS (Optionally)
Please send an E-Mail with your CV and transcript of records. Within the E-Mail please mention the related projects you have previously worked on.
E-Mail jonfrey@ethz.ch and cc tamiki@ethz.ch
Please send an E-Mail with your CV and transcript of records. Within the E-Mail please mention the related projects you have previously worked on.