NVIDIA ARDY streams text-driven 3D motion at real-time speed

Real-time controllable character motion has been a stubborn gap in the generative animation stack, so a new NVIDIA Research paper called ARDY that claims to bridge it is worth pausing on. The preprint on arXiv, by Kaifeng Zhao, Mathis Petrovich, Haotian Zhang, Tingwu Wang, Siyu Tang and Davis Rempe, was accepted to ACM Transactions on Graphics for SIGGRAPH 2026.

The framing in the abstract is the interesting part. Offline motion generators, the authors write, "offer precise control via text and kinematic constraints" but "lack the inference speed required for interactive settings," while online methods "enable real-time synthesis but often sacrifice controllability or struggle with complex text semantics and long-horizon goals." ARDY's pitch is a streaming generator that keeps both sides, taking online text prompts and flexible kinematic constraints and producing high-fidelity motion in real time.

The technical shape is a hybrid representation, explicit root features paired with a latent body embedding, feeding what the paper calls "a two-stage autoregressive transformer denoiser that features variable history context and supports conditioning on flexible, long-horizon kinematic constraints." It is trained on a large-scale motion capture dataset with ground-truth text labels and sampled constraints, so at inference it can take live prompting and long-horizon goals natively. Evaluation runs on HumanML3D plus a set the paper calls Bones Rigplay. An interactive demo, shown on the NVIDIA project page, features dynamic text control, keyframe pose constraints, path following, and locomotion driven by mouse and keyboard.

Why it matters: the abstract explicitly names animation, simulation, and humanoid robotics as the target applications, and the last of those is where the deployment bar has been highest. A motion prior that reacts to text and kinematic constraints without offline batching is exactly the kind of piece a robotics simulator or a game engine wants for character control.

The honest caveat is that the abstract does not publish inference latency numbers, hardware requirements, or head-to-head timings against prior text-to-motion baselines, and Bones Rigplay is described only as "large-scale, high-fidelity" without any licensing detail. Peer review at ACM TOG is a real signal of engineering rigor, but production reliability inside a live animation pipeline or a real humanoid is a different bar. Code and models are promised at the project page, and if the released artifacts hold up, animation and robotics teams get a controllable real-time motion generator they can actually plug in.