Master Thesis on “Reinforcement Learning and Graph Neural Networks for deformable object manipulation”

Supervisor: Prof. Ville Kyrki (ville.kyrki@aalto.fi)
Advisors: MSc. David Blanco Mulero (david.blancomulero@aalto.fi), MSc. Karol Arndt (karol.arndt@aalto.fi), PhD. Gökhan Alcan (gokhan.alcan@aalto.fi).

Keywords: reinforcement learning, graph neural networks, graph deep learning, robot learning and manipulation.

Project description

Reinforcement Learning (RL) has shown great results on tasks like learning to play games or dexterous manipulation. However, tasks like manipulation of deformable objects like cloth are still far from being solved. Recent approaches have imposed a graph structure using Graph Neural Networks (GNNs) on deformable objects as a means to better represent the underlying dynamics of the system.

This thesis will focus on how to combine RL methods with Graph Neural Networks for solving cloth manipulation tasks. The thesis will explore the state-of-the-art and work on solutions for cloth manipulation.

Previous knowledge on Machine Learning (MLPs, RNNs, LSTMs) is desired.

Deliverables

  • Review relevant RL and GNN literature,
  • Implementation of relevant RL+GNN algorithms,
  • Evaluation of the experimental results,
  • Evaluation of algorithms on real robots.

Practical information

Pre-requisites: Python (high/medium), Reinforcement Learning (medium), Machine Learning (medium)

Tools: OpenAI gym, Softgym/MuJoCo/PyBullet

Start: Available Immediately

References

Lin, Xingyu, Yufei Wang, and David Held. “Learning Visible Connectivity Dynamics for Cloth Smoothing.” arXiv preprint arXiv:2105.10389 (2021).

Wang, Tingwu, et al. “Nervenet: Learning structured policy with graph neural networks.” International Conference on Learning Representations. 2018.

Zambaldi, Vinicius, et al. “Deep reinforcement learning with relational inductive biases.” International Conference on Learning Representations. 2018.