ARDY: NVIDIA and ETH Zürich Introduce Real-Time Autoregressive Diffusion Model for Interactive 3D Human Motion

Researchers from NVIDIA and ETH Zürich have unveiled ARDY, an autoregressive diffusion framework for real-time interactive 3D human motion generation, accepted to ACM Transactions on Graphics (SIGGRAPH 2026). ARDY bridges the gap between offline motion generation systems—which offer precise control but lack interactive speed—and existing online methods that sacrifice controllability. It supports online text prompting and flexible long-horizon kinematic constraints including root trajectories, full-body keyframes, and sparse joint positions, all with real-time responsiveness.

ARDY's architecture centers on a hybrid motion representation combining explicit global root features with a compact latent body embedding, processed through a two-stage autoregressive transformer denoiser. Trained on a large-scale motion capture dataset, the model natively supports streaming generation suitable for animation, game character control, simulation, and humanoid robotics. The team also demonstrates integration with the SONIC physical tracking policy to enable interactive control of a Unitree G1 humanoid robot.