ASE

Why we included this project

Researchers working on physics-based character animation will find ASE a useful reference for the adversarial skill embedding method, which trains a reusable low-level controller to imitate a library of motion clips and then reuses that controller for downstream tasks such as heading control or getting up after a fall. The repo walks through a clear two-phase workflow, pre-training on motion datasets followed by task training, so it is a good way to see the full pipeline rather than a single model in isolation. One caveat worth knowing up front: the authors have deprecated this codebase and point to MimicKit for newer, easier-to-use implementations, so treat ASE as a research artifact for studying the approach and reproducing the paper rather than as maintained production software. It also requires NVIDIA Isaac Gym, so factor that into your environment setup.