GPC: A Generative Pretraining Framework for Transferable Physics-Based Character Animation
By
[Submitted on 28 Jun 2026]
Summary
This paper introduces Generative Pretrained Controllers (GPC), a framework for physics-based character animation that uses tokenization and next-token modeling to create reusable generative controllers from large-scale motion datasets. The approach combines end-to-end reinforcement learning with Finite Scalar Quantization (FSQ) to build a "motion vocabulary," and a GPT-style autoregressive transformer to model the structure of that vocabulary. The resulting controller achieves a 99.98% success rate in reproducing motion clips, exhibits natural emergent behaviors like perturbation response and fall recovery, and can be fine-tuned for new downstream tasks.
Source
Twitter / XGPC: A Generative Pretraining Framework for Transferable Physics-Based Character Animationarxiv.orgKey quotes
· 4 pulledDeveloping controllers capable of completing a wide range of tasks in a natural and life-like manner is a key challenge in enabling practical applications of physics-based character animation.
Our proposed framework greatly simplifies the training process compared to previous tokenized methods, and achieves a 99.98% success rate in reproducing a vast corpus of motion clips.
The generative controller exhibits a variety of natural emergent behaviors, such as responsive behaviors to perturbations and recovery behaviors after falling.
This results in highly robust general purpose controllers for a variety of downstream applications.
You might also wanna read
General Physics Transformer Achieves Foundation Model Capabilities for Multiple Physical Systems
Researchers present the General Physics Transformer (GPhyT), a physics foundation model trained on 1.8 TB of diverse simulation data that ca
arxiv.org·9mo ago
Lumos-Nexus: A Training-Efficient Two-Stage Framework for High-Fidelity Video Generation with Reasoning Capabilities
Lumos-Nexus is a training-efficient unified video generation framework that addresses the computational challenge of integrating large high-
arxiv.org·1mo ago
DynaFLIP: A Dynamics-Aware Multimodal Pre-Training Framework for Robot Manipulation Perception
DynaFLIP is a dynamics-aware multimodal pre-training framework for robot manipulation perception. It constructs image-language-3D flow tripl
arxiv.org·1mo ago
Using Curriculum Learning and PufferLib to Train Superhuman AI Agents for 2048 and Tetris
The article describes using PufferLib, a reinforcement learning framework, to train gaming agents that achieve superhuman performance in 204
ByteDance's Seedance 2.0 Shows Progress in AI Video Generation but Still Has Limitations
The article discusses ByteDance's Seedance 2.0, a new AI video generation model that shows impressive improvements in creating realistic vid
The Verge·4mo ago
Diffuse-CLoC: A Guided Diffusion Framework for Physics-Based Motion Control
Diffuse-CLoC is a guided diffusion framework designed for physics-based look-ahead control, enabling intuitive, steerable, and physically re
Comments
Sign in to join the conversation.
No comments yet. Be the first.