← ArchivePaper2022
ASE: Large-Scale Reusable Adversarial Skill Embeddings for Physically Simulated Characters
Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, Sanja Fidler
Large-scale adversarial skill embedding space for physically simulated characters enabling diverse and composable motion skill reuse.
Abstract
The incredible feats of athleticism demonstrated by humans are made possible in part by a vast repertoire of general-purpose motor skills, acquired through years of practice and experience. These skills not only enable humans to perform complex tasks, but also provide powerful priors for guiding their behaviors when learning new tasks. This is in stark contrast to what is common practice in physics-based character animation, where control policies are most typically trained from scratch for each task. In this work, we present a large-scale data-driven framework for learning versatile and reusable skill embeddings for physically simulated characters. Our approach combines techniques from adversarial imitation learning and unsupervised reinforcement learning to develop skill embeddings that produce life-like behaviors, while also providing an easy to control representation for use on new downstream tasks. Our models can be trained using large datasets of unstructured motion clips, without requiring any task-specific annotation or segmentation of the motion data. By leveraging a massively parallel GPU-based simulator, we are able to train skill embeddings using over a decade of simulated experiences, enabling our model to learn a rich and versatile repertoire of skills. We show that a single pre-trained model can be effectively applied to perform a diverse set of new tasks.
How to read this
- Category
- Method: large-scale reusable skill embeddings for physics-based character control
- Contributions
- A data-driven framework that learns versatile, reusable skill embeddings for physically simulated characters
- Combines adversarial imitation learning with unsupervised reinforcement learning to produce life-like, controllable behaviors
- Trains from large unstructured motion-clip datasets with no task-specific annotation or segmentation, leveraging massively parallel GPU simulation
- Context
- Extends adversarial-motion-prior control (AMP, Peng et al. 2021) from per-task imitation toward a pre-trained, composable skill latent space reused across downstream tasks.Builds on: AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control
- Correctness
- The premise is that an unsupervised skill space learned from unstructured clips transfers to new tasks; demonstrated in simulation with a GPU-parallel simulator, so reader caveats are sim-only results and dependence on motion-data coverage and reward/adversary tuning.
- Clarity
- Idea is graspable on a first pass; the adversarial plus unsupervised-RL objective and the embedding/encoder design need a slower read.
- How to read it
- First pass for the pretrain-skills-then-reuse idea and how it differs from AMP; second pass on the training objectives if you plan to build downstream controllers on top.
Related work
- CALM: Conditional Adversarial Latent Models for Directable Virtual Characters 2023 / SIGGRAPH
- AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control 2021 / SIGGRAPH
- C·ASE: Learning Conditional Adversarial Skill Embeddings for Physics-based Characters 2023 / SIGGRAPH Asia
- SuperPADL: Scaling Language-Directed Physics-Based Control with Progressive Supervised Distillation 2024 / SIGGRAPH
Keywords
This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →