MotionGPT is an open-source research model with just under 2,000 stars on GitHub that treats human motion as a foreign language, unifying motion generation and understanding inside a large language model — text in, motion out, and vice versa.
Presented at NeurIPS 2023, MotionGPT starts from an elegant idea: if motion sequences are tokenised the way words are, a language model can learn motion as another language it speaks. The resulting unified model handles both directions of the motion-language problem — generating human motion from a textual description and producing language about motion it is given — within one framework rather than separate task-specific networks. That matters for animation, games, robotics and embodied AI, anywhere instructions in words need to become movement in the world. The repository provides the official implementation from OpenMotionLab, including the motion tokenisation and model code behind the paper, and the project site showcases generated examples. Development continued into 2025. As a research release it expects a research workflow — datasets, checkpoints, GPU training — but for anyone working on text-to-motion, it remains one of the reference points the subfield cites.
MotionGPT is free and open source under the MIT license. You supply the compute and motion datasets; there is no commercial service or paid tier attached to the project.
Researchers in animation, robotics and embodied AI working on text-driven motion generation, and LLM engineers curious about extending language models to a genuinely non-textual modality.