- ARDY is an autoregressive diffusion model designed for interactive human motion generation, bridging the gap between offline quality and online real-time responsiveness.
- It supports online text prompting and flexible long-horizon kinematic constraints, including root paths/waypoints, full-body keyframes, and sparse joint positions/rotations.
- The model uses a hybrid motion representation combining explicit global root features with a compact latent body embedding, balancing precise trajectory control with efficient generative learning.
- Its two-stage autoregressive transformer denoiser features variable history context and conditions on sparse, long-horizon constraints, predicting root motion first and then body motion conditioned on it.
- ARDY enables interactive applications like game character control, robotics, and simulation, demonstrated with real-time locomotion control and integration with the SONIC policy on the Unitree G1 robot.