nv-tlabs/ardy

ARDY is a research project from NVIDIA that generates realistic human motion for characters in real time. It comes from a SIGGRAPH 2026 paper and lets a user control a moving character by typing text prompts and setting spatial goals, such as a path to walk along, specific poses to hit, or the position of certain joints, while the character keeps moving smoothly and responsively. This repository provides the code, pretrained models, and a demo so people can try the published models rather than train from scratch. Setting it up needs a fairly capable machine. It expects a recent NVIDIA GPU and CUDA setup, a Python 3.10 or newer environment, and PyTorch installed first so the right build matches your specific graphics card. Building the project also compiles a bundled C plus plus extension for correcting motion, which needs CMake and a C++17 compiler. The text understanding part of the model relies on a gated Meta Llama 3 model from Hugging Face, so a user needs a Hugging Face account with access granted to that model plus an access token before anything will run. Several pretrained checkpoints are listed, trained on different character skeletons, frame rates, and prediction lengths, and these download automatically the first time the demo uses them. The main way to try the project is an interactive demo that runs in a web browser. It shows the generated motion live, lets a user steer the character with the mouse or keyboard, place waypoints for it to walk toward, or type text prompts describing how it should move. There is also an optional dataset of motion sequences for testing the version of the demo that pulls constraints from real recorded movement rather than manual input. This project is aimed at researchers and engineers working on character animation, robotics, or generative motion models, not at casual users, since it requires GPU hardware, gated model access, and a fair amount of setup before the demo runs. The code is released under an Apache 2.0 license, though the pretrained model checkpoints themselves use a separate NVIDIA open model license.