Master Thesis on “Real2Sim Transfer for Cloth Manipulation”

Supervisor: Prof. Ville Kyrki (ville.kyrki@aalto.fi)

Advisors: Dr. Gökhan Alcan (gokhan.alcan@aalto.fi), David Blanco Mulero (david.blancomulero@aalto.fi),

Keywords: cloth manipulation, real2sim transfer, high-fidelity simulators.

Project Description

Deformable objects such as cloths have many degrees of freedom, which creates additional challenges while manipulating them. Utilization of a simulated environment in learning the policies for cloth manipulation is usually the preferred option. “High fidelity” simulations of clothes are pivotal components, especially if the policy is targeted to be transferred to a real-world application.

This thesis will focus on proposing principled ways to increase the fidelity of cloth simulations using real-world observations from different types/sizes of cloths. Observations can include RGB-D measurements as well as absolute positions and velocities retrieved from a motion capture system. The thesis will comparatively investigate the available simulation environments in terms of adjustable parameters and work on the best suitable one for Real2Sim transfer.

Deliverables

  • Review of relevant literature,
  • Comparisons of different cloth simulators in terms of adjustable parameters,
  • Principled ways of tuning cloth parameters with real-world observations,
  • Evaluations of the proposed method with different types + sizes of cloths,
  • Manipulation in real robot using policies learned in simulation.

Practical Information

Pre-requisites: Python(high/medium), ROS (medium)

Tools: Franka Panda Robot, OptiTrack Motion Capture System, Microsoft Kinect or Microsoft Azure Kinect

Simulators: (up-to-change) MuJoCo, PyBullet, SoftGym

Start: Available immediately

References