Master's Thesis: Mixture of Experts Fixture Learning
Virtual Fixtures are an important tool to support the operator in haptic teleoperation when dealing with complicated tasks in the remote environment. Static Virtual Fixtures are however often not flexible enough when visual perception is involved as targets might appear or disappear at any time with their position being updated in real time.
Newly developed Probabilistic Visual Servoing Fixtures based on a Mixture of Experts (MoE) formulation allow for a principled way to define and arbitrate between fixtures based on such perceptual input. Their formulation however currently requires manual tuning of hyperparameters which prevents non-expert users from using this method. The goal of this thesis is to develop a method for demonstration-guided programming of the fixture parameters.
Tasks
- Literature research on learning-based approaches for the Mixture of Experts gating function,
- integration of a vision-based tracking system for obtaining target poses,
- estimation of the MoE hyperparameters using Maximum Likelihood estimation,
- extension of the method by e.g. smoothing transitions, learning dead zones and parameter ranges,
- evaluation of the method.
Qualification
- Good knowledge about foundations of robotics (Rigid body transformations, kinematics, etc.)
- Good knowledge of Probabilistic Machine Learning methods
- Strong programming skills in Python 3 and C++
- Familiar with development on Linux operating systems
- Experience with Git, CI/CD pipelines