JeasLee's picture
Upload folder using huggingface_hub (#2)
d5d60d0

RoboInter-Data: LeRobot v2.1 Format (Actions + Annotations + Videos)

The primary data format of RoboInter-Data. Contains robot actions, camera observations, and rich intermediate representation annotations in LeRobot v2.1 format (parquet + MP4 videos), ready for policy training. Especially, we calculate the delta EEF (gripper) action of Droid (instead of the joint velocity or the origin cartesian action of the base).

Sub-dataset Source Robot Episodes Frames Tasks Image Size Raw Image Size
lerobot_droid_anno DROID Franka + Robotiq 152,986 46,259,014 43,026 320 x 180 640 x 360
lerobot_rh20t_anno RH20T Multiple 82,894 40,755,632 146 320 x 180 640 x 360

Both datasets share fps=10, chunks_size=1000, and the same annotation schema.

Directory Layout

lerobot_droid_anno/  (or lerobot_rh20t_anno/)
├── meta/
│   ├── info.json              # Dataset metadata (fps, features, shapes, etc.)
│   ├── episodes.jsonl         # Per-episode info (one JSON per line)
│   ├── episodes_stats.jsonl   # Per-episode statistics
│   └── tasks.jsonl            # Task/instruction mapping
├── data/
│   └── chunk-{NNN}/           # Parquet data chunks (1,000 episodes per chunk)
│       ├── episode_000000.parquet
│       ├── episode_000001.parquet
│       └── ...
└── videos/
    └── chunk-{NNN}/
        ├── observation.images.primary/
        │   └── episode_{NNNNNN}.mp4
        └── observation.images.wrist/
            └── episode_{NNNNNN}.mp4

Data Fields

Core Fields (Shared by DROID & RH20T)

Field Shape Type Description
action (7,) float64 Delta EEF action: [delta_x, delta_y, delta_z, delta_rx, delta_ry, delta_rz, gripper_command]
state (7,) float64 EEF state: [x, y, z, rx, ry, rz, gripper_state]
observation.images.primary (180, 320, 3) video (H.264) Primary camera RGB video
observation.images.wrist (180, 320, 3) video (H.264) Wrist camera RGB video

Metadata Fields (Shared)

Field Type Description
episode_name string Episode unique identifier, e.g. "3072_exterior_image_1_left"
camera_view string Camera perspective, e.g. "exterior_image_1_left"
task string Task language description (via task_index -> tasks.jsonl)
episode_index int64 Episode index in dataset
frame_index int64 Frame index within episode
timestamp float32 Timestamp in seconds (frame_index / fps)
index int64 Global frame index across all episodes
task_index int64 Index into tasks.jsonl

Other Information Fields — DROID Only

lerobot_droid_anno contains the following additional fields from the original DROID dataset:

Field Shape Type Description
other_information.language_instruction_2 (1,) string Alternative language instruction (source 2)
other_information.language_instruction_3 (1,) string Alternative language instruction (source 3)
other_information.action_delta_tcp_pose (7,) float64 Delta TCP pose action: [dx, dy, dz, drx, dry, drz, gripper]
other_information.action_delta_wrist_pose (7,) float64 Delta wrist pose action: [dx, dy, dz, drx, dry, drz, gripper]
other_information.action_tcp_pose (7,) float64 Absolute TCP pose: [x, y, z, rx, ry, rz, gripper]
other_information.action_wrist_pose (7,) float64 Absolute wrist pose: [x, y, z, rx, ry, rz, gripper]
other_information.action_gripper_velocity (1,) float64 Gripper velocity
other_information.action_joint_position (7,) float64 Joint position action: [j1..j7]
other_information.action_joint_velocity (7,) float64 Joint velocity action: [j1..j7]
other_information.action_cartesian_velocity (6,) float64 Cartesian velocity: [vx, vy, vz, wx, wy, wz]
other_information.observation_joint_position (7,) float64 Observed joint positions: [j1..j7]
other_information.observation_gripper_position (1,) float64 Observed gripper position
other_information.observation_gripper_open_state (1,) float64 Gripper open state
other_information.observation_gripper_pose6d (6,) float64 Gripper 6D pose: [x, y, z, rx, ry, rz]
other_information.observation_tcp_pose6d (6,) float64 TCP 6D pose: [x, y, z, rx, ry, rz]
other_information.is_first (1,) bool First frame flag
other_information.is_last (1,) bool Last frame flag
other_information.is_terminal (1,) bool Terminal state flag

Other Information Fields — RH20T Only

lerobot_rh20t_anno contains the following additional fields from the original RH20T dataset:

Field Shape Type Description
other_information.action_delta_tcp_pose (7,) float64 Delta TCP pose action: [dx, dy, dz, drx, dry, drz, gripper]
other_information.action_tcp_pose (7,) float64 Absolute TCP pose: [x, y, z, rx, ry, rz, gripper]
other_information.gripper_command (1,) float64 Gripper command
other_information.observation_joint_position (14,) float64 Observed joint positions: [j1..j14]
other_information.observation_gripper_open_state (1,) float64 Gripper open state
other_information.observation_gripper_pose6d (6,) float64 Gripper 6D pose: [x, y, z, rx, ry, rz]
other_information.tcp_camera (7,) float64 TCP in camera frame: [x, y, z, qx, qy, qz, qw]
other_information.tcp_base (7,) float64 TCP in base frame: [x, y, z, qx, qy, qz, qw]
other_information.gripper (1,) string Gripper metadata (JSON)
other_information.is_first (1,) bool First frame flag
other_information.is_last (1,) bool Last frame flag
other_information.is_terminal (1,) bool Terminal state flag

Key difference: DROID has 7-DoF joint positions and richer action representations (wrist pose, joint/cartesian velocities). RH20T has 14-DoF joint positions, TCP transforms in camera/base frames, and gripper metadata JSON.


Annotation Fields (Shared by DROID & RH20T)

All annotation fields are prefixed with annotation. and stored as JSON strings. Empty string "" means no annotation is available for that frame.

Field Format Description
annotation.time_clip [[start, end], ...] Subtask temporal segments (frame ranges)
annotation.instruction_add string Structured task language instruction
annotation.substask string Current subtask description
annotation.primitive_skill string Primitive skill label (pick, place, push, twist, etc.)
annotation.segmentation string Segmentation reference (path)
annotation.object_box [[x1, y1], [x2, y2]] Manipulated object bounding box
annotation.placement_proposal [[x1, y1], [x2, y2]] Target placement bounding box
annotation.trace [[x, y], ...] Future 10-frame gripper trajectory waypoints
annotation.gripper_box [[x1, y1], [x2, y2]] Gripper bounding box
annotation.contact_frame int / -1 Frame index when gripper contacts object (-1 = past contact)
annotation.state_affordance [x, y, z, rx, ry, rz] 6D EEF state at contact frame
annotation.affordance_box [[x1, y1], [x2, y2]] Gripper bounding box at contact frame
annotation.contact_points [x, y] Contact point in pixel coordinates
annotation.origin_shape [h, w] Original image resolution for coordinate reference

Bounding Box Format

All bounding boxes use pixel coordinates with origin at top-left:

[[x1, y1], [x2, y2]]   // [top-left, bottom-right]

Trace Format

10 future waypoints for gripper trajectory prediction:

[[110, 66], [112, 68], [115, 70], [118, 72], [120, 75], [122, 78], [125, 80], [128, 82], [130, 85], [132, 88]]

Q_Annotation Fields (Quality Indicators, Shared)

Each annotation has a corresponding quality indicator prefixed with Q_annotation.:

Field Values Description
Q_annotation.instruction_add "Primary" / "Secondary" / "" Instruction quality
Q_annotation.substask "Primary" / "Secondary" / "" Subtask quality
Q_annotation.primitive_skill "Primary" / "Secondary" / "" Primitive skill quality
Q_annotation.segmentation "Primary" / "Secondary" / "" Segmentation quality
Q_annotation.object_box "Primary" / "Secondary" / "" Object box quality
Q_annotation.placement_proposal "Primary" / "Secondary" / "" Placement proposal quality
Q_annotation.trace "Primary" / "Secondary" / "" Trace quality
Q_annotation.gripper_box "Primary" / "Secondary" / "" Gripper box quality
Q_annotation.contact_frame "Primary" / "Secondary" / "" Contact frame quality
Q_annotation.state_affordance "Primary" / "Secondary" / "" State affordance quality
Q_annotation.affordance_box "Primary" / "Secondary" / "" Affordance box quality
Q_annotation.contact_points "Primary" / "Secondary" / "" Contact points quality
  • Primary: High-confidence annotation
  • Secondary: Acceptable quality, may have minor errors
  • "" (empty): No annotation available

Quick Start

The dataloader code is at RoboInterData/lerobot_dataloader.

Installation

pip install numpy torch pyarrow av opencv-python

Basic Usage

from lerobot_dataloader import create_dataloader

dataloader = create_dataloader(
    "path/to/lerobot_droid_anno",
    batch_size=32,
    action_horizon=16,
)

for batch in dataloader:
    images = batch["observation.images.primary"]  # (B, H, W, 3)
    actions = batch["action"]                     # (B, 16, 7)
    trace = batch["annotation.trace"]             # Parsed JSON lists
    skill = batch["annotation.primitive_skill"]   # List of strings
    break

Multiple Datasets (DROID + RH20T)

dataloader = create_dataloader(
    [
        "path/to/lerobot_droid_anno",
        "path/to/lerobot_rh20t_anno",
    ],
    batch_size=32,
    action_horizon=16,
)

for batch in dataloader:
    print(batch["dataset_name"])  # Source dataset identifier
    break

Data Filtering

Frame Range Filtering

Remove idle frames at episode start/end using range_nop.json:

dataloader = create_dataloader(
    "path/to/lerobot_droid_anno",
    range_nop_path="path/to/range_nop.json",
)

Format of range_nop.json:

{
  "3072_exterior_image_1_left": [12, 217, 206]
}

[start_frame, end_frame, valid_length] — frames outside this range are idle/stationary.

Q_Annotation Filtering

Select episodes by annotation quality:

from lerobot_dataloader import create_dataloader, QAnnotationFilter

# Only Primary quality
dataloader = create_dataloader(
    "path/to/lerobot_droid_anno",
    q_filters=[
        QAnnotationFilter("Q_annotation.instruction_add", ["Primary"]),
        QAnnotationFilter("Q_annotation.gripper_box", ["Primary"]),
    ]
)

# Any non-empty annotation
dataloader = create_dataloader(
    "path/to/lerobot_droid_anno",
    q_filters=[
        QAnnotationFilter("Q_annotation.trace", ["not_empty"])
    ]
)

Combined Filtering

from lerobot_dataloader import FilterConfig, QAnnotationFilter

config = FilterConfig(
    range_nop_path="path/to/range_nop.json",
    q_filters=[
        QAnnotationFilter("Q_annotation.trace", ["Primary", "Secondary"]),
    ],
    q_filter_mode="all",  # "all" = AND, "any" = OR
)

dataloader = create_dataloader("path/to/lerobot_droid_anno", filter_config=config)

Transforms

from lerobot_dataloader import Compose, Normalize, ResizeImages, ToTensorImages, LeRobotDataset
from lerobot_dataloader.transforms import compute_stats

# Compute normalization stats
dataset = LeRobotDataset("path/to/lerobot_droid_anno", load_videos=False)
stats = compute_stats(dataset)

# Create transform pipeline
transform = Compose([
    ResizeImages(height=224, width=224),
    ToTensorImages(),   # (H,W,C) uint8 -> (C,H,W) float32
    Normalize(stats),
])

dataloader = create_dataloader("path/to/lerobot_droid_anno", transform=transform)

Direct Dataset Access

from lerobot_dataloader import LeRobotDataset
from lerobot_dataloader.transforms import ParseAnnotations

dataset = LeRobotDataset(
    "path/to/lerobot_droid_anno",
    transform=ParseAnnotations(),
)

print(f"Total frames: {len(dataset)}")
print(f"Total episodes: {dataset.num_episodes}")
print(f"FPS: {dataset.fps}")

sample = dataset[0]
print(f"Action: {sample['action']}")
print(f"Object box: {sample['annotation.object_box']}")
print(f"Skill: {sample['annotation.primitive_skill']}")

Format Conversion

The LeRobot v2.1 format was converted from original data + LMDB annotations using:


Related Resources

Resource Link
RoboInter-Data (parent dataset) HuggingFace
RoboInter Project GitHub
DataLoader Code lerobot_dataloader
Conversion Scripts convert_to_lerobot
Demo Visualizer RoboInterData-Demo
DROID Dataset droid-dataset.github.io
RH20T Dataset rh20t.github.io

License

Please refer to the original dataset licenses for RoboInter, DROID, and RH20T.