Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Leveraging and enforcing symmetries for deep RL motion policies
In this thesis we aim to leverage symmetries in the morphology and physics of robotic systems for deep reinforcement learning. Projections into subspaces, special parametrizations and physics-informed neural networks could thereby increase sample efficiency, but more importantly directly enforce symmetric behavior in the learned motion policies. Symmetric behavior is sometimes hard to achieve through loss functions only, but is often highly sought-after, as it can lead to better controllability, safety and overall practicability of the robot.
The scope of this project suggests a semester thesis, but if there is interest the scope could be extended to a master thesis.
Related work:
[1] Learning to walk in minutes using massively parallel deep reinforcement learning, https://arxiv.org/abs/2109.11978
[2] On the Continuity of Rotation Representations in Neural Networks, https://arxiv.org/abs/1812.07035
The scope of this project suggests a semester thesis, but if there is interest the scope could be extended to a master thesis.
Related work:
[1] Learning to walk in minutes using massively parallel deep reinforcement learning, https://arxiv.org/abs/2109.11978
[2] On the Continuity of Rotation Representations in Neural Networks, https://arxiv.org/abs/1812.07035
- Literature Review
- Precise problem formulation
- Formation of ideas and derivation of related theory
- Implementation thereof in a RL framework with Python
- Experimental evaluation in simulation and potentially the real world
- Literature Review - Precise problem formulation - Formation of ideas and derivation of related theory - Implementation thereof in a RL framework with Python - Experimental evaluation in simulation and potentially the real world
- Theoretical basis in RL / ML
- Solid basis in maths and/or dynamics
- Experience in Python
- Experience in C++ is a plus
- Theoretical basis in RL / ML - Solid basis in maths and/or dynamics - Experience in Python - Experience in C++ is a plus
Victor Klemm (vklemm@ethz.ch)
Nikita Rudin (rudinn@ethz.ch)
Victor Klemm (vklemm@ethz.ch) Nikita Rudin (rudinn@ethz.ch)
