← ArchivePaper2023
C·ASE: Learning Conditional Adversarial Skill Embeddings for Physics-based Characters
Zhiyang Dou, Xuelin Chen, Qingnan Fan, Taku Komura, Wenping Wang
Divides heterogeneous motion into homogeneous subsets for training a conditional adversarial model that controls diverse skill behaviors.
Abstract
We present C · ASE, an efficient and effective framework that learns Conditional Adversarial Skill Embeddings for physics-based characters. C · ASE enables the physically simulated character to learn a diverse repertoire of skills while providing controllability in the form of direct manipulation of the skills to be performed. This is achieved by dividing the heterogeneous skill motions into distinct subsets containing homogeneous samples for training a low-level conditional model to learn the conditional behavior distribution. The skill-conditioned imitation learning naturally offers explicit control over the character’s skills after training. The training course incorporates the focal skill sampling, skeletal residual forces, and element-wise feature masking to balance diverse skills of varying complexities, mitigate dynamics mismatch to master agile motions and capture more general behavior characteristics, respectively. Once trained, the conditional model can produce highly diverse and realistic skills, outperforming state-of-the-art models, and can be repurposed in various downstream tasks. In particular, the explicit skill control handle allows a high-level policy or a user to direct the character with desired skill specifications, which we demonstrate is advantageous for interactive character animation.
How to read this
- Category
- Method: conditional adversarial skill embedding for physics-based character control
- Contributions
- Learns conditional adversarial skill embeddings (C-ASE) so a simulated character acquires a diverse skill repertoire with direct, explicit skill control
- Divides heterogeneous skill motions into homogeneous subsets to train a low-level conditional model of the skill behavior distribution
- Introduces focal skill sampling, skeletal residual forces, and element-wise feature masking to balance varied skills, ease dynamics mismatch, and capture general behavior
- Context
- Extends the adversarial skill embedding line of work, directly building on ASE (Peng et al. 2022) by adding explicit skill conditioning.Builds on: ASE: Large-Scale Reusable Adversarial Skill Embeddings for Physically Simulated Characters
- Correctness
- Claims to outperform prior models and to be repurposable downstream, but gains rest on the assumption that heterogeneous motions can be cleanly partitioned into homogeneous subsets, so reusability depends on the quality of that division.
- Clarity
- Fairly accessible at a high level; a first pass conveys the conditioning idea, a second pass is needed for the sampling, residual-force, and masking mechanics.
- How to read it
- Focus on how the skill subsets are formed and how the conditioning handle is exposed to a high-level policy; a second pass pays off for the three training mechanisms (focal sampling, residual forces, feature masking).
Builds on
Built upon by
Nothing yet.
Related work
- CALM: Conditional Adversarial Latent Models for Directable Virtual Characters 2023 / SIGGRAPH
- DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills 2018 / SIGGRAPH
- Character Controllers Using Motion VAEs 2020 / SIGGRAPH
- DReCon: Data-Driven Responsive Control of Physics-Based Characters 2019 / TOG
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 →