# MotionGPT: Human Motion as a Foreign Language

**Authors:** Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, Tao Chen  
**arXiv:** [2306.14795](https://arxiv.org/abs/2306.14795)  
**Date:** 2023-06-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.14795v2/

## Key points

- MotionGPT is a motion-language model designed to handle various motion-related tasks.
- It combines language data with large-scale motion models for enhanced performance.
- The model uses discrete vector quantization to represent human motion as "motion tokens".
- MotionGPT is pre-trained on a mixture of motion-language data and fine-tuned on prompt-based question-and-answer tasks.
- It achieves state-of-the-art performance in text-driven motion generation, captioning, prediction, and interpolation tasks.
- Comparisons with other models highlight MotionGPT's superiority in terms of performance.
- Treating human motion as a foreign language opens up new possibilities for understanding and generating humanlike motions.

## Summary

MotionGPT is a unified and versatile motion-language model that aims to handle various motion-related tasks. It combines language data with large-scale motion models to enable motion-language pre-training that enhances the performance of motion-related tasks. The model employs discrete vector quantization for human motion and transfers 3D motion into "motion tokens", similar to word tokens in text generation. This "motion vocabulary" allows for unified language modeling on both text and motion data, treating human motion as a specific language. Inspired by prompt learning, MotionGPT is pre-trained with a mixture of motion-language data and fine-tuned on prompt-based question-and-answer tasks. Extensive experiments demonstrate that MotionGPT achieves state-of-the-art performance on multiple motion tasks including text-driven motion generation, captioning, prediction, and interpolation. Additionally, the paper provides comparisons with other existing models such as T2M-GPT, MLD, TM2T, MDM, and MotionDiffuse to highlight its superiority in terms of performance. The proposed model opens up new possibilities for understanding and generating humanlike motions by treating them as a foreign language.