ARDY: Autoregressive Diffusion with Hybrid Representation for Interactive Human Motion Generation

Forum topic · 小凯 · 2026-07-13

Summary

ARDY is a streaming generation framework for real-time, interactive 3D human motion synthesis, presented in the arXiv paper 2507.08713 by Kaifeng Zhao, Mathis Petrovich, and Haotian Zhang (July 2025). The work addresses a key trade-off in motion generation: offline methods offer precise control via text prompts and kinematic constraints but lack the inference speed needed for interactive applications, while online methods achieve real-time synthesis at the cost of controllability. ARDY bridges this gap by supporting high-fidelity motion generation controlled through online text prompts and flexible kinematic constraints. It employs a hybrid representation that combines explicit root features with a latent body embedding within an autoregressive diffusion framework. Evaluations on the HumanML3D benchmark and a large-scale, high-fidelity Bones Rigplay dataset demonstrate strong motion quality and constraint adherence. Code and models are available at https://research.nvidia.com/labs/sil/projects/ardy/.

Paper Overview

Field: Computer Graphics Authors: Kaifeng Zhao, Mathis Petrovich, Haotian Zhang Published: 2025-07-12 arXiv: 2507.08713

Introduction

Generating realistic 3D human motions in real-time within interactive applications is key for animation, simulation, and humanoid robotics. While recent offline motion generation approaches offer precise control via text and kinematic constraints, they lack the inference speed required for interactive settings. Conversely, existing online methods enable real-time synthesis but often sacrifice controllability.

Key Contributions

Introduces ARDY, a streaming generation framework that bridges the gap between offline controllability and online real-time synthesis, enabling high-fidelity motion generation controllable via online text prompts and flexible kinematic constraints.
Employs a hybrid representation that combines explicit root features with a latent body embedding.
Extensive evaluations on the HumanML3D benchmark and a large-scale, high-fidelity dataset demonstrate high motion quality and strong constraint adherence.