Recent generative animation models can create realistic character motion rather than simply replaying existing animation clips. In this project, you will use Epic Games’ new AnimGen framework and motion-capture data from the large-scale HiPHI dataset to investigate human-object interaction.
The student will select a manageable subset of interactions, such as approaching, reaching for, picking up or moving objects, prepare this data and train an AnimGen character controller. The main research question will be whether the trained model can generalise when the position or orientation of an object is changed. Generated animations will be evaluated for interaction accuracy and realism, potentially including a perceptual experiment.
The project involves generative AI, motion capture, character animation and Unreal Engine 5. Students should be taking Real-Time Animation and a Machine Learning module, have strong programming skills, and will require access to a suitable GPU for model training.