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

2026-07-09

AI Summary

Existing approaches struggle to simultaneously satisfy the demands of real-time performance, controllability, and long-horizon semantic understanding in interactive 3D human motion generation. This work proposes ARDY, a novel framework that uniquely integrates autoregressive diffusion mechanisms with a hybrid motion representation—combining explicit root trajectory features and implicit body embeddings—to support variable-length historical context and flexible kinematic constraints. Leveraging a two-stage autoregressive Transformer denoiser conditioned on text labels and user-specified constraints derived from real pose samples, ARDY enables high-fidelity, online text-driven motion synthesis. Experiments on the HumanML3D and Bones Rigplay datasets demonstrate its superior generation quality and strong adherence to complex constraints, while an interactive demo showcases practical capabilities such as dynamic text control, keyframe editing, and path following.

Abstract

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 or struggle with complex text semantics and long-horizon goals due to limited context windows. In this work, we introduce ARDY, a streaming generation framework that bridges this gap by enabling high-fidelity motion generation controllable via online text prompts and flexible kinematic constraints. ARDY employs a hybrid representation that combines explicit root features with a latent body embedding, balancing precise trajectory control with efficient generative learning. We propose a two-stage autoregressive transformer denoiser that features variable history context and supports conditioning on flexible, long-horizon kinematic constraints. By training on a large-scale motion capture dataset and being directly conditioned on text labels and kinematic constraints sampled from ground truth poses, ARDY natively learns controllable generation that supports online prompting and flexible long-horizon goals. Extensive evaluations on the HumanML3D benchmark and the large-scale, high-fidelity Bones Rigplay dataset demonstrate ARDY's high motion quality and constraint adherence, validating the efficacy of our key architectural decisions. Finally, we demonstrate the method's practical versatility through an interactive demo featuring dynamic text control, diverse keyframe pose constraints, path following, and interactive locomotion control via mouse and keyboard. Supplementary video results, code, and model releases can be found at https://research.nvidia.com/labs/sil/projects/ardy/.

Problem

• interactive human motion generation
• real-time motion synthesis
• controllability
• long-horizon goals
• text-to-motion

Innovation

• autoregressive diffusion
• hybrid representation
• interactive motion generation
• kinematic constraints
• real-time human animation