Teaching simulated humans to swordfight and get back up

What it does ASE is a two-level reinforcement-learning system for physics-based character animation built on NVIDIA Isaac Gym. It first pre-trains a low-level controller on motion-capture datasets to learn adversarial skill embeddings—reusable movement primitives like getting up after a fall or swinging a sword. These embeddings then act as a motor foundation for training high-level controllers to execute specific tasks such as reaching a target, striking, or maintaining a heading.

The interesting bit Instead of learning every task from scratch, the low-level controller absorbs a broad library of humanoid skills from motion clips, and the high-level controller merely orchestrates them. The repository also includes an implementation of Adversarial Motion Priors and pre-trained checkpoints for sword-and-shield tasks.

Key highlights

• Two-stage training: reusable low-level skill embeddings, then task-specific high-level control
• Includes pre-trained models and Reallusion sword-and-shield motion-capture data (noncommercial use only)
• Implements both ASE and the related Adversarial Motion Priors method
• Supports motion retargeting via ase/poselib/retarget_motion.py
• Built on NVIDIA Isaac Gym for physics simulation

Caveats

• The codebase is explicitly deprecated; the authors direct users to MimicKit for current implementations
• Depends on Isaac Gym, which requires a separate download from NVIDIA
• Bundled motion data is strictly noncommercial

Verdict Researchers studying hierarchical RL and physics-based animation will find the architecture instructive, but new projects should use MimicKit instead. As a deprecated reference implementation, it is more useful for understanding the paper than for production use.