RoboInter-Data: Intermediate Representation Annotations for Robot Manipulation

Rich, dense, per-frame intermediate representation annotations for robot manipulation, built on top of DROID and RH20T. Developed as part of the RoboInter project. You can try our Online demo.

The annotations cover 230k episodes and include: subtasks, primitive skills, segmentation, gripper/object bounding boxes, placement proposals, affordance boxes, grasp poses, traces, contact points, etc. And each with a quality rating (Primary / Secondary).

Dataset Structure

RoboInter-Data/
│
├── Annotation_with_action_lerobotv21/    # [Main] LeRobot v2.1 format (actions + annotations + videos)
│   ├── lerobot_droid_anno/               #   DROID: 152,986 episodes
│   └── lerobot_rh20t_anno/              #   RH20T:  82,894 episodes
│
├── Annotation_pure/                      # Annotation-only LMDB (no actions/videos)
│   └── annotations/                      #   35 GB, all 235,920 episodes
│
├── Annotation_raw/                       # Original unprocessed annotations
│   ├── droid_annotation.pkl              #   Raw DROID annotations (~20 GB)
│   ├── rh20t_annotation.pkl              #   Raw RH20T annotations (~11 GB)
│   └── segmentation_npz.zip.*            #   Segmentation masks (~50 GB, split archives)
│
├── Annotation_demo_app/                  # Small demo subset for online visualization
│   ├── demo_data/                        #   LMDB annotations for 20 sampled videos
│   └── videos/                           #   20 MP4 videos
│
├── Annotation_demo_larger/               # Larger demo subset for local visualization
│   ├── demo_annotations/                 #   LMDB annotations for 120 videos
│   └── videos/                           #   120 MP4 videos
│
├── All_Keys_of_Primary.json              # Episode names where all annotations are Primary quality
├── RoboInter_Data_Qsheet.json            # Per-episode quality ratings for each annotation type
├── RoboInter_Data_Qsheet_value_stats.json# Distribution statistics of quality ratings
├── RoboInter_Data_RawPath_Qmapping.json  # Mapping: original data source path -> episode splits & quality
├── range_nop.json                        # Non-idle frame ranges for all 230k episodes
├── range_nop_droid_all.json              # Non-idle frame ranges (DROID only)
├── range_nop_rh20t_all.json              # Non-idle frame ranges (RH20T only)
├── val_video.json                        # Validation set: 7,246 episode names
└── VideoID_2_SegmentationNPZ.json        # Episode video ID -> segmentation NPZ file path mapping

---

1. Annotation_with_action_lerobotv21 (Recommended)

The primary data format. Contains actions + observations + annotations in LeRobot v2.1 format (parquet + MP4 videos), ready for policy training.

Directory Layout

lerobot_droid_anno/  (or lerobot_rh20t_anno/)
├── meta/
│   ├── info.json              # Dataset metadata (fps=10, features, etc.)
│   ├── episodes.jsonl         # Episode information
│   └── tasks.jsonl            # Task/instruction mapping
├── data/
│   └── chunk-{NNN}/           # Parquet files (1,000 episodes per chunk)
│       └── episode_{NNNNNN}.parquet
└── videos/
    └── chunk-{NNN}/
        ├── observation.images.primary/
        │   └── episode_{NNNNNN}.mp4
        └── observation.images.wrist/
            └── episode_{NNNNNN}.mp4

Data Fields

Category	Field	Shape / Type	Description
Core	action	(7,) float64	Delta EEF: [dx, dy, dz, drx, dry, drz, gripper]
	state	(7,) float64	EEF state: [x, y, z, rx, ry, rz, gripper]
	observation.images.primary	(180, 320, 3) video	Primary camera RGB
	observation.images.wrist	(180, 320, 3) video	Wrist camera RGB
Annotation	annotation.instruction_add	string	Structured task language instruction
	annotation.substask	string	Current subtask description
	annotation.primitive_skill	string	Primitive skill label (pick, place, push, ...)
	annotation.object_box	JSON [[x1,y1],[x2,y2]]	Manipulated object bounding box
	annotation.gripper_box	JSON [[x1,y1],[x2,y2]]	Gripper bounding box
	annotation.trace	JSON [[x,y], ...]	Future 10-step gripper trajectory
	annotation.contact_frame	JSON int	Frame index when gripper contacts object
	annotation.contact_points	JSON [x, y]	Contact point pixel coordinates
	annotation.affordance_box	JSON [[x1,y1],[x2,y2]]	Gripper box at contact frame
	annotation.state_affordance	JSON [x,y,z,rx,ry,rz]	6D EEF state at contact frame
	annotation.placement_proposal	JSON [[x1,y1],[x2,y2]]	Target placement bounding box
	annotation.time_clip	JSON [[s,e], ...]	Subtask temporal segments
Quality	Q_annotation.*	string	Quality rating: "Primary" / "Secondary" / ""

Quick Start

The dataloader is located at our RoboInter Codebase.

from lerobot_dataloader import create_dataloader

# Single dataset
dataloader = create_dataloader(
    "path/to/Annotation_with_action_lerobotv21/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"]              # JSON strings
    skill = batch["annotation.primitive_skill"]    # List[str]
    break

# Multiple datasets (DROID + RH20T)
dataloader = create_dataloader(
    [
        "path/to/lerobot_droid_anno",
        "path/to/lerobot_rh20t_anno",
    ],
    batch_size=32,
    action_horizon=16,
)

Filtering by Quality & Frame Range

from lerobot_dataloader import create_dataloader, QAnnotationFilter

dataloader = create_dataloader(
    "path/to/lerobot_droid_anno",
    batch_size=32,
    range_nop_path="path/to/range_nop.json",       # Remove idle frames
    q_filters=[
        QAnnotationFilter("Q_annotation.trace", ["Primary"]),
        QAnnotationFilter("Q_annotation.gripper_box", ["Primary", "Secondary"]),
    ],
)

For full dataloader documentation and transforms, see: RoboInterData/lerobot_dataloader.

Format Conversion Scripts

The LeRobot v2.1 data was converted using:

• DROID: convert_droid_to_lerobot_anno_fast.py
• RH20T: convert_rh20t_to_lerobot_anno_fast.py

---

2. Annotation_pure (Annotation-Only LMDB)

Contains only the intermediate representation annotations (no action data, no videos) stored as a single LMDB database. Useful for lightweight access to annotations or as input for the LeRobot conversion pipeline. The format conversion scripts and corresponding lightweight dataloader functions are provided in lmdb_tool. You can downloade high-resolution videos by following Droid hr_video_reader and RH20T API.

Data Format

Each LMDB key is an episode name (e.g., "3072_exterior_image_1_left"). The value is a dict mapping frame indices to per-frame annotation dicts:

{
    0: {  # frame_id
        "time_clip": [[0, 132], [132, 197], [198, 224]],   # subtask segments
        "instruction_add": "pick up the red cup",           # language instruction
        "substask": "reach for the cup",                    # current subtask
        "primitive_skill": "reach",                         # skill label
        "segmentation": None,                               # (stored separately in Annotation_raw)
        "object_box": [[45, 30], [120, 95]],                # manipulated object bbox
        "placement_proposal": [[150, 80], [220, 140]],      # target placement bbox
        "trace": [[x, y], ...],                             # next 10 gripper waypoints
        "gripper_box": [[60, 50], [100, 80]],               # gripper bbox
        "contact_frame": 101,                               # contact event frame (−1 if past contact)
        "state_affordance": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6],# 6D EEF state at contact
        "affordance_box": [[62, 48], [98, 82]],             # gripper bbox at contact frame
        "contact_points": [[75, 65], [85, 65]],             # contact pixel coordinates
        ...
    },
    1: { ... },
    ...
}

Reading LMDB

import lmdb
import pickle

lmdb_path = "Annotation_pure/annotations"
env = lmdb.open(lmdb_path, readonly=True, lock=False, readahead=False)

with env.begin() as txn:
    # List all episode keys
    cursor = txn.cursor()
    for key, value in cursor:
        episode_name = key.decode("utf-8")
        episode_data = pickle.loads(value)

        # Access frame 0
        frame_0 = episode_data[0]
        print(f"{episode_name}: {frame_0['instruction_add']}")
        print(f"  object_box: {frame_0['object_box']}")
        print(f"  trace: {frame_0['trace'][:3]}...")  # first 3 waypoints
        break

env.close()

CLI Inspection Tool

cd RoboInter/RoboInterData/lmdb_tool

# Basic info
python read_lmdb.py --lmdb_path Annotation_pure/annotations --action info

# View a specific episode
python read_lmdb.py --lmdb_path Annotation_pure/annotations --action item --key "3072_exterior_image_1_left"

# Field coverage statistics
python read_lmdb.py --lmdb_path Annotation_pure/annotations --action stats --key "3072_exterior_image_1_left"

# Multi-episode summary
python read_lmdb.py --lmdb_path Annotation_pure/annotations --action summary --limit 100

---

3. Annotation_raw (Original Annotations)

The original, unprocessed annotation files before conversion to LMDB format. These files are large and slow to load.

File	Size	Description
droid_annotation.pkl	~20 GB	Raw DROID intermediate representation annotations
rh20t_annotation.pkl	~11 GB	Raw RH20T intermediate representation annotations
segmentation_npz.zip.*	~50 GB	Object segmentation masks (split archives)

Reading Raw PKL

cd /RoboInter-Data/Annotation_raw
cat segmentation_npz.zip.* > segmentation_npz.zip
unzip segmentation_npz.zip

import pickle

with open("Annotation_raw/droid_annotation.pkl", "rb") as f:
    droid_data = pickle.load(f)  # Warning: ~20 GB, takes several minutes

# droid_data[episode_key] contains raw intermediate representation data
# including: all_language, all_gripper_box, all_grounding_box, all_contact_point, all_traj, etc.

To convert raw PKL to the LMDB format used in Annotation_pure, see the conversion script in the RoboInter repository.

---

4. Demo Subsets (Annotation_demo_app & Annotation_demo_larger)

Pre-packaged subsets for quick visualization using the RoboInterData-Demo Gradio app. Both subsets share the same LMDB annotation format + MP4 video structure.

Subset	Videos	Size	Use Case
Annotation_demo_app	20	~929 MB	HuggingFace Spaces online demo
Annotation_demo_larger	120	~12 GB	Local visualization with more examples

Running the Visualizer

git clone https://github.com/InternRobotics/RoboInter.git
cd RoboInter/RoboInterData-Demo

# Option A: Use the small demo subset (for Spaces)
ln -s /path/to/Annotation_demo_app/demo_data ./demo_data
ln -s /path/to/Annotation_demo_app/videos ./videos

# Option B: Use the larger demo subset (for local)
ln -s /path/to/Annotation_demo_larger/demo_annotations ./demo_data
ln -s /path/to/Annotation_demo_larger/videos ./videos

pip install -r requirements.txt
python app.py
# Open http://localhost:7860

The visualizer supports all annotation types: object segmentation masks, gripper/object/affordance bounding boxes, trajectory traces, contact points, grasp poses, and language annotations (instructions, subtasks, primitive skills).

---

5. Metadata JSON Files

Quality & Filtering

File	Description
All_Keys_of_Primary.json	List of 65,515 episode names where all annotation types are rated Primary quality.
RoboInter_Data_Qsheet.json	Per-episode quality ratings for every annotation type. Each entry contains Q_instruction_add, Q_substask, Q_trace, etc. with values "Primary", "Secondary", or null.
RoboInter_Data_Qsheet_value_stats.json	Distribution of quality ratings across all episodes.
RoboInter_Data_RawPath_Qmapping.json	Mapping from original data source paths to episode splits and their quality ratings.

Frame Ranges (Idle Frame Removal)

File	Description
range_nop.json	Non-idle frame ranges for all 235,920 episodes (DROID + RH20T).
range_nop_droid_all.json	Non-idle frame ranges for DROID episodes only.
range_nop_rh20t_all.json	Non-idle frame ranges for RH20T episodes only.

Format: { "episode_name": [start_frame, end_frame, valid_length] }

import json

with open("range_nop.json") as f:
    range_nop = json.load(f)

# Example: "3072_exterior_image_1_left": [12, 217, 206]
# Means: valid action frames are 12~217, total 206 valid frames
# (frames 0~11 and 218+ are idle/stationary)

Other

File	Description
val_video.json	List of 7,246 episode names reserved for the validation set.
VideoID_2_SegmentationNPZ.json	Mapping from episode video ID to the corresponding segmentation NPZ file path in Annotation_raw/segmentation_npz. null if no segmentation is available.

---

Related Resources

Resource	Link
Project	RoboInter
VQA Dataset	RoboInter-VQA
VLM Checkpoints	RoboInter-VLM
LMDB Tool	RoboInterData/lmdb_tool
High-Resolution Video Reader	RoboInterData/hr_video_reader
LeRobot DataLoader	RoboInterData/lerobot_dataloader
LeRobot Conversion	RoboInterData/convert_to_lerobot
Demo Visualizer	RoboInterData-Demo
Online Demo	HuggingFace Space
Raw DROID Dataset	droid-dataset.github.io
Raw RH20T Dataset	rh20t.github.io

License

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