GPC: Large-Scale Generative Pretraining for Transferable Motor Control

This paper proposes Generative Pretrained Controllers (GPC), a new way to build physics-based character controllers by treating motion as a token prediction problem. Instead of training a separate policy for each behavior, the authors first learn a discrete “motion vocabulary” with Finite Scalar Quantization (FSQ), then train a GPT-style autoregressive transformer over those codes to generate controls.

The practical appeal is reuse. Once the codebook and controller are trained, the system can be finetuned for downstream tasks, which could reduce the amount of bespoke RL work needed for each new animation behavior. The authors say the framework simplifies training compared with earlier tokenized methods and can reproduce a large corpus of motion clips with 99.98% success.

For game devs, the most relevant part is the quality of the emergent behavior: the controller reportedly reacts to perturbations and can recover after falling, which is exactly the kind of robustness that matters for gameplay-driven animation. That suggests a path toward more general-purpose character controllers that can survive messy runtime conditions instead of only looking good in curated clips.

The paper is from Yi Shi, Yifeng Jiang, Chen Tessler, and Xue Bin Peng, submitted to arXiv on June 28, 2026, and listed with a SIGGRAPH 2026 proceedings reference. It sits at the intersection of graphics, ML, and physics-based animation, so the near-term audience is mainly researchers and technical animation teams, but the longer-term implication is broader: fewer one-off controllers, more adaptable...