Apppublic

cs686/ardy-motion-api

sourceHugging Faceupdated 2mo agoView on Hugging Face

6likes

App README

ARDY Motion Playground

Interactive text-to-motion demo for ARDY (Autoregressive Diffusion with Hybrid Representation for Interactive Human Motion Generation, NVIDIA, SIGGRAPH 2026).

Pick a rig (human or robot), type a natural-language prompt, generate a 3D motion clip, then orbit, scrub, and play it in the built-in Three.js playground (play/pause, timeline scrubber, speed control, loop, and a root-trajectory trail).

—Human rig: nvidia/ARDY-Core-RP-20FPS-Horizon40 — 27-joint "Core" skeleton at 20 fps.
—Robot rig: nvidia/ARDY-G1-RP-25FPS-Horizon52 — 34-joint Unitree G1 humanoid robot at 25 fps (same rig as the sibling Space hugging-apps/ardy-g1-motion-generation).
—Text encoder: LLM2Vec-Llama-3-8B using the public unsloth/llama-3-8b-Instruct base plus the original McGill MNTP and supervised adapters. No Meta-gated base-model access is required.
—Hardware: ZeroGPU.

Post-processing (foot-skate cleanup, which needs a compiled C++ extension) is disabled in this demo; generation uses the model's autoregressive path directly.

Based on the nv-tlabs/ardy reference implementation (scripts/generate.py).

Blender export

Open Blender export to generate three downloadable files:

—ardy_motion.bvh — directly importable with Blender's BVH importer.
—ardy_motion.npz — complete ARDY output including joint positions, local and global rotations, root trajectory, and foot contacts.
—metadata.json — prompt, seed, FPS, skeleton hierarchy, and coordinate-system information.

The programmatic endpoint is /generate_blender. The existing /generate endpoint and Three.js playground remain available.

The base and adapter repositories can be overridden with the Space variables LLM2VEC_BASE_MODEL, LLM2VEC_MNTP_MODEL, and LLM2VEC_SUPERVISED_MODEL.