DRL-based physics modeling of aged locomotion for motion simulation, stability analysis and data augmentation

Locomotion is critical to human mobility, allowing people to freely move in the physical world. Locomotion modeling has been one major topic in character animation research, where intelligent virtual avatars (IVAs) are trained to perform realistic motions in VR/AR and game applications. However, there is very little research effort dedicated to modeling movements for any demographic group, nor for health concerns and mobility studies.

This project is the first work to learn physics models of age-specific locomotion through Deep Reinforcement Learning (DRL), for both motion simulation and mobility analysis purposes. One benefit of having intelligent physics models is that simulations and experiments can be performed through IVAs in the virtual world, without involving real humans in inaccessible or dangerous environments in reality. The goal of this project is to learn intelligent physics models of locomotion for older adults, based on which we can conduct various mobility experiments, and generate a large synthetic locomotion dataset, through IVAs’ simulated motions in the virtual environment.

First, ground-truth locomotion from older adults of various age groups will be collected through mocap and provided as learning examples. Physics models will be trained using the state-of-the-art DRL method, to find optimal policies for IVAs to exert forces and torques to walk and run like the ground truth. Once trained, push-recovery and surface stability experiments will be conducted in the simulated environment, and report results from different age groups to understand the human aging process. All IVAs with their trained physics models will be put into various virtual scenarios to simulate realistic locomotion for creating a large synthetic dataset.

This project will lead to the publication of novel approaches for training age-specific physics models and findings from stability experiments. This project will also release a large synthetic dataset, which could benefit health researchers, computer scientists and animation artists.