Master Thesis on “Dynamic Cloth Manipulation Using Reinforcement Learning”

Supervisor: Prof. Ville Kyrki (ville.kyrki@aalto.fi)
Advisor: Dr. Gökhan Alcan (gokhan.alcan@aalto.fi)

Keywords: simulation, robotics, manipulation, sim-to-real transfer, reinforcement learning.

Project Description

Deformable objects such as cloths have many degrees of freedom, which creates additional challenges while manipulating them compared to rigid objects. The forces caused by the acceleration define the manipulation task as static or dynamic. Even though the dynamic cloth manipulation is more challenging due to the difficulties in the characterization of high-dimensional configurations spaces, it generates some advantages by allowing to control the non-grasped points of the cloth as well.

The goal of this thesis is to develop reinforcement learning-based approaches for dynamic cloth manipulation. In this context, the thesis is expected to include high-fidelity cloth simulations, training of an agent in that simulation environment to perform dynamic manipulation tasks, and transferring the learned policies to a real robot in testing its generalization capabilities with different sizes of clothes.

Deliverables

  • Related state-of-the-art literature review,
  • Realization of high-fidelity cloth simulations in MuJoCo environment (domain randomization),
  • Development of a reinforcement-learning based control method in simulation to perform dynamic manipulation tasks, e.g., sideways folding,
  • Transfer of the learned policies to a real robot (Franka Emika) and test the policy with different sizes and materials of cloths (sim-to-real).

Practical Information

Pre-requisites: Reinforcement Learning, Robotics, Computer Vision.
Tools: Python, PyTorch/Tensorflow, MuJoCo, ROS
Platform: Franka Emika
Start: Available immediately

References