seed

BONES-SEED: Skeletal Everyday Embodiment Dataset BONES-SEED is an open dataset of 142,220 annotated human motion animations for humanoid robotics. It provides motion capture data in SOMA and Unitree G1 formats, with natural language descriptions, temporal segmentation, and detailed skeletal metadata. Project website: bones.studio/datasets/seed Interactive viewer: seed-viewer.bones.studio Associated code: github.com/bones-studio/seed-viewer Total motions 142,220 (71,132 original + 71,088 mirrored) Total duration ~288 hours (@ 120 fps) Performers 522 actors (253 F / 269 M) Age range 17–71 years Height range 145–199 cm Weight range 38–145 kg Output formats SOMA Uniform · SOMA Proportional · Unitree G1 MuJoCo-compatible Annotation depth Up to 6 NL descriptions per motion + temporal segmentation + technical descriptions + skeletal metadata Intended Uses BONES-SEED is designed to support research and development in: Humanoid whole-body control — training language-conditioned policies for humanoid robots Motion generation — text-to-motion and action-to-motion synthesis Motion retrieval — natural language search over large motion libraries Sim-to-real transfer — leveraging MuJoCo-compatible G1 trajectories for simulation training Imitation learning — learning from diverse human demonstrations Motion understanding — temporal segmentation, style classification, and activity recognition Download BONES-SEED is hosted on Hugging Face and can be downloaded using any of the methods below. Using the Hugging Face Hub Browse and download files directly from the dataset repository: https://huggingface.co/datasets/bones-studio/seed Using Git LFS Make sure Git LFS is installed git lfs install Clone the full dataset git clone https://huggingface.co/datasets/bones-studio/seed Using the Hugging Face CLI Install the Hugging Face CLI if you haven't already pip install huggingface_hub Download the full dataset huggingface-cli download bones-studio/seed --repo-type dataset --local-dir ./bones-