BONES-SEED (Skeletal Everyday Embodiment Dataset) 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-seed

Using Python

from huggingface_hub import snapshot_download

# Download the full dataset
snapshot_download(
    repo_id="bones-studio/seed",
    repo_type="dataset",
    local_dir="./bones-seed"
)

Loading Metadata Only

import pandas as pd

# Load directly from Hugging Face
df = pd.read_parquet(
    "hf://datasets/bones-studio/seed/metadata/seed_metadata_v002.parquet"
)
print(f"Total motions: {len(df)}")
print(f"Columns: {df.columns.tolist()}")

Dataset Structure

After downloading and extracting, the dataset is organized as follows:

bones-seed/
├── metadata/
│   ├── seed_metadata_v003.parquet          # Main metadata (51 columns × 142,220 rows)
│   ├── seed_metadata_v003.csv              # Same metadata in CSV format
│   └── seed_metadata_v002_temporal_labels.jsonl  # Temporal segmentation labels
├── soma_uniform/
│   └── bvh/{date}/{motion_name}.bvh        # SOMA Uniform motion files
├── soma_proportional/
│   └── bvh/{date}/{motion_name}.bvh        # SOMA Proportional motion files
├── g1/
│   └── csv/{date}/{motion_name}.csv        # Unitree G1 MuJoCo-compatible joint trajectories
├── soma_shapes/
│   ├── soma_base_fit_mhr_params.npz        # Shared shape params (SOMA Uniform)
│   ├── soma_proportion_fit_mhr_params/
│   │   └── {actor_id}.npz                  # Per-actor shape params (SOMA Proportional)
│   └── soma_base_rig/
│       ├── soma_base_skel_minimal.bvh      # SOMA base skeleton definition (BVH)
│       └── soma_base_skel_minimal.usd      # SOMA base skeleton definition (USD)
└── LICENSE.md

Unpacking

The motion data directories (soma_uniform/, soma_proportional/, g1/) are distributed as tar archives. After downloading, extract them into the dataset root:

tar -xf soma_uniform.tar
tar -xf soma_proportional.tar
tar -xf g1.tar

Motion Categories

BONES-SEED spans a wide range of human activities organized into 8 top-level packages and 20 fine-grained categories.

Packages

Package	Motions	Description
Locomotion	74,488	Walking, jogging, jumping, climbing, crawling, turning, and transitions
Communication	21,493	Gestures, pointing, looking, and communicative body language
Interactions	14,643	Object manipulation, pick-and-place, carrying, and tool use
Dances	11,006	Full-body dance performances across multiple styles
Gaming	8,700	Game-inspired actions and dynamic movements
Everyday	5,816	Household tasks, consuming, sitting, reading, and daily activities
Sport	3,993	Athletic movements and sports-specific actions
Other	2,081	Stunts, martial arts, magic, and edge-case motions

Data Formats

Every motion is provided in three skeletal representations supporting two character models: SOMA and Unitree G1 robot. SOMA is a canonical body topology and rig that acts as a universal pivot for parametric human body models.

SOMA Proportional (BVH)

A per-actor skeleton that preserves the original performer's body proportions. Each actor has an individual shape file.

soma_proportional/bvh/{date}/{motion_name}.bvh
soma_shapes/soma_proportion_fit_mhr_params/{actor_id}.npz

SOMA Uniform (BVH)

A standardized skeleton shared across all motions, enabling direct comparison and batch processing. Each motion file is paired with a single shared shape file. The base skeleton definition is provided in both BVH and USD formats.

soma_uniform/bvh/{date}/{motion_name}.bvh
soma_shapes/soma_base_fit_mhr_params.npz
soma_shapes/soma_base_rig/soma_base_skel_minimal.bvh
soma_shapes/soma_base_rig/soma_base_skel_minimal.usd

Unitree G1 MuJoCo-compatible (CSV)

Joint-angle trajectories retargeted to the Unitree G1 humanoid robot.

g1/csv/{date}/{motion_name}.csv

Annotations

Each motion in BONES-SEED comes with rich multimodal annotations designed for language-conditioned policy learning, motion retrieval, and motion generation.

Natural Language Descriptions

Every motion includes up to 6 natural language descriptions at varying levels of detail:

Natural descriptions (4): Fluent, human-written descriptions from different perspectives
Technical description (1): Precise biomechanical description of the motion
Short descriptions (2): Concise labels for indexing and retrieval

Example — read_newspaper_sitting:

Field	Text
`content_natural_desc_1`	character reading newspaper while sitting
`content_natural_desc_2`	person reads a newspaper while sitting
`content_natural_desc_3`	individual sits and reads a newspaper
`content_natural_desc_4`	A person sitting reads a newspaper, holding it with both hands, moving pages and folding the newspaper.
`content_technical_description`	reading a newspaper holding it with both hands while sitting, moving pages folding a newspaper
`content_short_description`	reading newspaper sitting

Temporal Segmentation Labels

Each motion includes temporal segmentation that breaks the full sequence into meaningful phases with precise timestamps and natural language descriptions. These labels were created by NVIDIA for the Kimodo project and are stored in metadata/seed_metadata_v002_temporal_labels.jsonl (one JSON object per line).

Schema:

Field	Type	Description
`filename`	string	Motion filename (matches `filename` column in metadata)
`num_events`	int	Number of temporal segments
`events`	array	Ordered list of temporal segments
`events[].start_time`	float	Segment start time in seconds
`events[].end_time`	float	Segment end time in seconds
`events[].description`	string	Natural language description of the segment

Example — inside_door_knob_left_side_open_R_002__A512:

{
  "filename": "inside_door_knob_left_side_open_R_002__A512",
  "num_events": 3,
  "events": [
    {"start_time": 0.0, "end_time": 1.88, "description": "A person rotates the door knob with their right hand."},
    {"start_time": 1.88, "end_time": 3.53, "description": "A person opens the door outward from the inside, holding the knob and then lowers their hand."},
    {"start_time": 3.53, "end_time": 4.83, "description": "A person is standing idle and slightly moving their right hand."}
  ]
}

Loading temporal labels:

import json

temporal_labels = {}
with open("metadata/seed_metadata_v002_temporal_labels.jsonl") as f:
    for line in f:
        entry = json.loads(line)
        temporal_labels[entry["filename"]] = entry["events"]

# Look up segments for a specific motion
events = temporal_labels["inside_door_knob_left_side_open_R_002__A512"]
for event in events:
    print(f"[{event['start_time']:.2f}s - {event['end_time']:.2f}s] {event['description']}")

Motion Properties

Each motion is tagged with structured metadata for filtering and analysis:

Field	Description	Example Values
`content_type_of_movement`	Primary movement type	walking, jogging, gesture, dancing, jumping
`content_body_position`	Starting/primary body position	standing, sitting on floor, crouching, crawling
`content_uniform_style`	Performance style	neutral, injured leg, injured torso, hurry, old
`content_horizontal_move`	Horizontal displacement flag	0 or 1
`content_vertical_move`	Vertical displacement flag	0 or 1
`content_props`	Involves props/objects	0 or object descriptor
`content_complex_action`	Multi-phase complex action	0 or 1
`content_repeated_action`	Contains repeated cycles	0 or 1

Metadata Schema

The metadata parquet file contains 51 columns organized into five groups.

Motion Identity

Column	Type	Description
`move_name`	string	Unique motion identifier
`filename`	string	Base filename (without extension)
`move_duration_frames`	int	Duration in frames (@ 120 fps)
`package`	string	Top-level category (Locomotion, Communication, etc.)
`category`	string	Fine-grained category
`is_neutral`	float	Whether the motion uses a neutral performance style
`is_mirror`	bool	Whether the motion is a left-right mirror

File Paths

Column	Type	Description
`move_soma_uniform_path`	string	Path to SOMA Uniform BVH file
`move_soma_uniform_shape_path`	string	Path to SOMA Uniform shape parameters
`move_soma_proportional_path`	string	Path to SOMA Proportional BVH file
`move_soma_proportional_shape_path`	string	Path to SOMA Proportional shape parameters
`move_g1_mujoco_path`	string	Path to Unitree G1 MuJoCo-compatible CSV file

Capture Session

Column	Type	Description
`take_name`	string	Capture session identifier
`take_actor`	string	Actor identifier for this take
`take_org_name`	string	Original take name
`take_date`	int	Capture date (YYMMDD format)
`take_day_part`	string	Part of capture day

Content Annotations

Column	Type	Description
`content_name`	string	Semantic motion name
`content_natural_desc_1`	string	Natural language description 1
`content_natural_desc_2`	string	Natural language description 2
`content_natural_desc_3`	string	Natural language description 3
`content_natural_desc_4`	string	Natural language description 4
`content_technical_description`	string	Technical/biomechanical description
`content_short_description`	string	Short description 1
`content_short_description_2`	string	Short description 2
`content_all_rigplay_styles`	string	All performance styles applied
`content_uniform_style`	string	Normalized style label
`content_type_of_movement`	string	Movement type classification
`content_body_position`	string	Body position classification
`content_horizontal_move`	int	Horizontal displacement flag
`content_vertical_move`	int	Vertical displacement flag
`content_props`	string	Props/objects involved
`content_complex_action`	int	Complex action flag
`content_repeated_action`	int	Repeated action flag

Actor Biometrics

Column	Type	Description
`actor_uid`	string	Unique actor identifier
`actor_height`	string	Height category (S / M / T)
`actor_height_cm`	int	Height in centimeters
`actor_foot_cm`	int	Foot length in cm
`actor_collarbone_height_cm`	int	Collarbone height in cm
`actor_collarbone_span_cm`	int	Collarbone span in cm
`actor_elbow_span_cm`	int	Elbow span in cm
`actor_wrist_span_cm`	int	Wrist span in cm
`actor_shoulder_span_cm`	int	Shoulder span in cm
`actor_hips_height_cm`	int	Hips height in cm
`actor_hips_bones_span_cm`	int	Hips bone span in cm
`actor_knee_height_cm`	int	Knee height in cm
`actor_ankle_height_cm`	int	Ankle height in cm
`actor_weight_kg`	int	Weight in kilograms
`actor_age_yr`	int	Age in years
`actor_gender`	string	Gender (F / M)
`actor_profession`	string	Performer background (actor, dancer, stuntman, general, professional)

About Bones Studio

With over 5 years of experience, Bones Studio builds enterprise-grade, multimodal datasets of human behavior and motion for AI and robotics. BONES-SEED represents a curated subset of Bones Studio's broader motion capture library, with expanded datasets available for commercial licensing.

Learn more: bones.studio/datasets

Acknowledgments

Thanks to NVIDIA for providing the SOMA and G1 retargets, and for creating the temporal segmentation labels as part of the Kimodo project.

Downloads last month: 15

Category	Motions
Basic Locomotion Neutral	33,430
Baseline	22,878
Gestures	17,590
Object Manipulation	11,620
Dancing	11,006
Object Interaction	10,817
Basic Locomotion Styles	10,746
Advanced Locomotion	6,036
Sports	3,973
Communication	3,723
Unusual Locomotion	3,242
Other	2,081
Consuming	1,388
Household	1,318
Stunts	858
Environments	614
Complex Actions	540
Looking and Pointing	180
Magic	160
Martial Arts	20