ARDY: NVIDIA Open Real Time Text to Motion Model for Digital Humans and Robots

NVIDIA Research published ARDY on July 10, 2026, an autoregressive diffusion model that generates 3D human and humanoid motion from text prompts in real time. The model streams output frame by frame, supports kinematic constraints at inference time, and was accepted to ACM Transactions on Graphics for SIGGRAPH 2026. Inference code is published under Apache 2.0; model weights are available under the NVIDIA Open Model License, which explicitly permits commercial use.

What ARDY Does

ARDY takes a text prompt and generates a 3D motion sequence one frame at a time. The autoregressive approach lets a production tool begin receiving and previewing motion data while the model is still computing the rest of the sequence, rather than waiting for a completed batch output. The model was built at NVIDIA Research's Spatial Intelligence Lab, with contributions from researchers at ETH Zurich.

Key Capabilities

The four core capabilities cover the main production requirements for text driven character animation: generating motion indefinitely, streaming in real time, applying spatial constraints, and adapting to body configurations beyond the human norm.

Infinite generation means the model does not require a defined sequence length at the start of a generation. A character can be directed through a scene with successive prompts, and the model continues generating motion as long as input is provided. Text streaming allows the model to begin generating motion before the full prompt is completed, enabling a conversational prompting approach.

Constrained generation allows a text prompt to be combined with geometric specifications that steer the output spatially. Robot motion generation is handled by the G1-RP variant, which was trained on the Unitree G1 configuration and produces motion at 25fps matched to that hardware.

Online Text to Motion Generation

Online generation refers to ARDY's capacity to stream motion frame by frame as it is computed. The model does not require a complete pass over the sequence before producing output. A character begins moving as soon as the first frames are generated, with subsequent frames appended in real time.

The demo above shows a character walking with a limp, generated directly from a text description. The motion reflects the specific quality described in the prompt rather than defaulting to a generic walk cycle. For animation workflows, this is the difference between a placeholder and a first pass that already contains the performance note specified by a director.

NVIDIA Research describes the approach as an autoregressive diffusion process that conditions each generated frame on previous frames, maintaining motion continuity across what would otherwise be a series of independent generation steps.

Kinematically Constrained Motion Generation

Kinematic constraints let a user specify where a character should be in space, not only what the character should be doing. A root trajectory defines a path through the scene. End effector targets specify where hands or feet should land. Full body keyframes fix the character's entire pose at designated frames.

The constraint system operates at inference time, meaning no retraining is required to specify new spatial requirements. This differs from models that require constraint information to be baked into training data or addressed through separate fine tuning. For production pipelines, constraints can be added, modified, or removed per shot without touching the model.

Four Model Variants

ARDY ships with four checkpoints covering two body configurations and two horizon lengths.

Application: Interactive Humanoid Control

ARDY can be driven interactively, with a user providing spatial targets through mouse clicks or keyboard input rather than prompts prepared in advance. The model responds in real time, generating motion that navigates a character through a scene toward positions the user defines.

The interactive control demos run in a browser viewport. A user clicks a target location in the scene, and the character navigates there with motion generated in real time. The system handles both spatial navigation and motion style simultaneously, producing different gaits depending on the text description provided alongside the spatial target.

Keyboard control generates character motion in response to directional input. Scene navigation shows a character moving through an environment with obstacle awareness. Both demos run in a browser, which is one indicator of the model's inference speed. ARDY generates motion quickly enough to be responsive in an interactive web application.

Humanoid Robot Control

The G1-RP variant generates motion for the Unitree G1 humanoid robot. NVIDIA trained this variant on robot motion data to match the G1's joint limits and locomotion characteristics. The output addresses what the physical hardware can execute, not just what looks plausible in a 3D viewport.

For film and animation production, robot motion generation is relevant in two contexts. Productions featuring humanoid robots as characters can use G1-RP to generate motion that reflects the mechanical constraints of real robot locomotion. As humanoid robots become more common in virtual production environments, G1-RP also provides a way to simulate their behavior in previz before a physical robot is present on set.

Filmmaking and Virtual Production Applications

ARDY's primary value for film production is generating performance motion from text without a motion capture session. A director can describe how a character should move and receive 3D motion data that can be applied to a rig in Unreal Engine or Maya without booking a performer or a stage.

The real time streaming capability is particularly relevant for previsualization. Shot blocking typically requires either a live performer or an animator spending time on each shot. ARDY can generate a first pass of character motion for an entire sequence from shot descriptions alone, giving a director something to respond to before committing production resources.

Character motion generated from text covers a different part of the production workflow than motion transfer. Wan2.2-Animate's character motion transfer applies a recorded source performance to a target character image. ARDY generates the motion directly from a text description. The two approaches cover different scenarios: ARDY for building motion from scratch when no recorded reference exists, Wan2.2-Animate for applying an existing performance to a new character.

NVIDIA Cosmos 3 provides physics aware video generation for full scenes, while ARDY addresses character motion specifically. Productions that need physically consistent environments and character behavior can use both: Cosmos 3 for world generation and ARDY for character movement within those environments.

For AI generated video production workflows, AI FILMS Studio's video workspace supports text-to-video and image-to-video generation with character-focused and scene generation models available alongside each other.

License and Model Access

Model weights for all four ARDY variants are available on HuggingFace under the NVIDIA Open Model License Agreement. The license explicitly permits commercial use and allows teams to create and distribute derivative models, provided attribution is maintained. This is distinct from Apache 2.0 or CC-BY licensing. The NVIDIA Open Model License has its own terms that govern redistribution and commercial deployment, and teams should review those terms directly before building production pipelines.

The inference code is published under Apache 2.0 on GitHub (nv-tlabs/ardy). Apache 2.0 covers the code without restriction. Commercial deployment of model weights requires review of the NVIDIA Open Model License specifically.

The model card on HuggingFace notes that "NVIDIA has verified this model meets prescribed quality standards," which is NVIDIA Research's standard language for releases from its inference and generation groups. This distinguishes the release from community contributed checkpoints.

Teams building pipelines around ARDY should test on representative samples of their specific motion requirements before committing. The HumanML3D benchmark results provide a general quality reference, but performance on any given production's motion requirements depends on how closely those requirements match the Bones Rigplay training distribution. Evaluating the model on motion categories that match your content is the appropriate verification step before production use.

---

Sources

arXiv: arxiv.org/abs/2607.08741 GitHub: nv-tlabs/ardy HuggingFace: nvidia/ardy collection Project page: research.nvidia.com/labs/sil/projects/ardy/