ASE

The incredible feats of athleticism demonstrated by humans are made possible in part a vast repertoire general-purpose motor skills, acquired through years practice and experience. These skills not only enable to perform complex tasks, but also provide powerful priors for guiding their behaviors when learning new tasks. This is stark contrast what common physics-based character animation, where control policies most typically trained from scratch each task. In this work, we present large-scale data-driven framework versatile reusable skill embeddings physically simulated characters. Our approach combines techniques adversarial imitation unsupervised reinforcement develop that produce life-like behaviors, while providing an easy representation use on downstream models can be using large datasets unstructured motion clips, without requiring any task-specific annotation or segmentation the data. By leveraging massively parallel GPU-based simulator, able train over decade experiences, enabling our model learn rich skills. We show single pre-trained effectively applied diverse set system allows users specify tasks simple reward functions, embedding then enables automatically synthesize naturalistic strategies order achieve task objectives.