Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Learning to Navigate in Rough Terrain with Deep Reinforcement Learning
This project aims to develop navigation policies suitable for locomotion in rough terrain using deep reinforcement learning.
Keywords: Robotics, Deep Reinforcement Learning, Autonomous Navigation
Navigation is a key aspect of autonomous mobile systems such as our quadrupedal robot ANYmal [1]. In the coming years, the RSL will take part in several competitions where ANYmal will have to navigate autonomously in harsh environments over long time horizons. A typical objective will be to navigate from a point A to B given a point cloud or a height map of the location.
Deep reinforcement learning is a promising method to learn complex behaviors for legged systems, but it is still limited to short time horizons [2]. The goal of this project is to extend current reinforcement learning based techniques and to show that it is possible to find navigation policies that fulfill user-defined objectives over longer horizons. For example, an objective could be to minimize the energy consumption, or to maximize the probability of success when going from A to B.
One possible way to tackle this challenge is to design a cascade controller: a first neural network outputs the target pose given a height map and the robot’s current state. This information is then sent to a pose-tracking controller. The pipeline can be trained using typical reinforcement learning algorithms such as TRPO [3].
[1] https://ieeexplore.ieee.org/document/7758092/
[2] https://arxiv.org/pdf/1707.02286.pdf
[3] https://arxiv.org/abs/1502.05477
Navigation is a key aspect of autonomous mobile systems such as our quadrupedal robot ANYmal [1]. In the coming years, the RSL will take part in several competitions where ANYmal will have to navigate autonomously in harsh environments over long time horizons. A typical objective will be to navigate from a point A to B given a point cloud or a height map of the location.
Deep reinforcement learning is a promising method to learn complex behaviors for legged systems, but it is still limited to short time horizons [2]. The goal of this project is to extend current reinforcement learning based techniques and to show that it is possible to find navigation policies that fulfill user-defined objectives over longer horizons. For example, an objective could be to minimize the energy consumption, or to maximize the probability of success when going from A to B.
One possible way to tackle this challenge is to design a cascade controller: a first neural network outputs the target pose given a height map and the robot’s current state. This information is then sent to a pose-tracking controller. The pipeline can be trained using typical reinforcement learning algorithms such as TRPO [3].
[1] https://ieeexplore.ieee.org/document/7758092/
[2] https://arxiv.org/pdf/1707.02286.pdf
[3] https://arxiv.org/abs/1502.05477
- Thorough literature research
- Develop navigation policies given a height map/point cloud of the world
- Deploy the pipeline on the robot and show that it works with a real life demo
- Thorough literature research - Develop navigation policies given a height map/point cloud of the world - Deploy the pipeline on the robot and show that it works with a real life demo
- Hands-on experience in deep learning
- Excellent working knowledge of C++
- Experience with TensorFlow is a plus
- Hands-on experience in deep learning - Excellent working knowledge of C++ - Experience with TensorFlow is a plus