Our group is part of the Agents and Simulated Reality department at the DFKI. We focus on application oriented research related to the realistic simulation of human characters. Human motion synthesis finds use in different applications ranging from work place simulations and the simulation of traffic scenarios to games.
Currently we work on the following research areas:
Deep Learning based Motion Synthesis
Neural networks are general function approximators and have shown great results at learning complex human motion from example data. We investigate different network architectures to learn motion models with large variation of styles.
Semantic Motion Segmentation
The raw motion capture data needs to be further processed and annotated to be usable in applications. We investigate different methods for reducing manual effort in the data processing using machine learning.
Physics-based Motion Synthesis
Physics simulations enable the realistic interaction of human avatars with their environment. We investigate the training of physics controllers using sampling-based optimization and reinforcement learning based on example motion data.
Motion Style Transfer
The capturing of large motion repertoires is expensive. Therefore it is desiable to be able to transfer the style of one motion to another. We investigate different methods to modify the style of large datasets based on a few example clips.
Statistical Motion Synthesis
Classical machine learning methods based on statistical modeling are an efficient way of creating reusable motion models. In past projects we have applied statistical modelling of human motion to practical applications and have explored different data representations.