NVIDIA researchers present ARDY, a streaming generation framework designed to enable high-fidelity 3D human motion synthesis in real-time interactive applications. The system bridges the gap between offline precision and online speed by combining explicit root features with latent body embeddings.
- Utilizes a hybrid representation that balances precise trajectory control with efficient generative learning.
- Employs a two-stage autoregressive transformer denoiser capable of variable history context.
- Supports conditioning on flexible, long-horizon kinematic constraints and online text prompts.
- Trained on large-scale motion capture data to natively learn controllable generation.
- Validated on HumanML3D and Bones Rigplay benchmarks for motion quality and constraint adherence.
The framework allows users to control motion via dynamic text, keyframe poses, path following, and interactive locomotion, making it suitable for animation, simulation, and humanoid robotics.