HorizonRobotics/Real-World-Dataset

Latest Updates

• [2025-10-23] v0.3 Dataset Released. 304 episodes across five tasks. This dataset was collected using two separate evaluation robot arms (one in Beijing and one in Shanghai). This can be understood as data from systems that have slight differences in their kinematic or sensor calibration parameters.
• [2025-10-21] We have released the real-bot mcap data packaging tool (supporting both Arrow and LMDB formats). This tool can package mcap files recorded by the Data Recorder Toolkit. Packer Source Code ref to Code, Usage Instructions ref to Docs
• [2025-10-14] v0.2 & v0.1 Dataset Updated: fixed the depth decode issue of the v0.2 dataset and added visualization for the v0.2 dataset.
• [2025-09-24] v0.2 Dataset Released: 1015 episodes across five tasks. Available in both Arrow and LMDB formats. The v0.2 dataset was created after we re-calibrated the zero point of our robotic arm.
• [2025-09-04] v0.1 Dataset Released: 1086 episodes across five tasks. Available in Arrow and LMDB formats. (See note on zero-point drift).

1. Data Introduction

Data Format

This project provides robotic manipulation datasets in two formats: Arrow and LMDB:

• Arrow Dataset: Built on the Apache Arrow format. Its column-oriented structure offers flexibility and will be the primary format for development in robo_orchard_lab. It features standardized message types and supports exporting to Mcap files for visualization.
• LMDB Dataset: Built on the LMDB (Lightning Memory-Mapped Database) format, which is optimized for extremely fast read speeds.

⚠ Important Note on Dataset Versions

The v0.1 dataset was affected by a robotic arm zero-point drift issue during data acquisition. We have since re-calibrated the arm and collected the v0.2 dataset.

• v0.2: Please use this version for all fine-tuning and evaluation to ensure model accuracy.
• v0.1: This version should only be used for pre-training experiments or deprecated entirely.

Verifying Hardware Consistency

If you are using your own Piper robot arm, you can check for the same zero-point drift issue:

1. Check Hardware Zero Alignment: Home the robot arm and visually inspect if each joint aligns correctly with the physical zero-point markers.
2. Replay v0.2 Dataset: Replay the joint states from the v0.2 dataset. If the arm successfully completes the tasks, your hardware setup is consistent with ours.

1.1 Version 0.3

Task	Episode Num	LMDB Dataset	Arrow Dataset
place_shoe	30	lmdb_dataset_place_shoe_2025_10_21_bj	arrow_dataset_place_shoe_2025_10_21_bj
	31	lmdb_dataset_place_shoe_2025_10_21_sh	arrow_dataset_place_shoe_2025_10_21_sh
empty_cup_place	30	lmdb_dataset_empty_cup_place_2025_10_21_bj	arrow_dataset_empty_cup_place_2025_10_21_bj
	31	lmdb_dataset_empty_cup_place_2025_10_21_sh	arrow_dataset_empty_cup_place_2025_10_21_sh
put_bottles_dustbin	30	lmdb_dataset_put_bottles_dustbin_2025_10_21_bj	arrow_dataset_put_bottles_dustbin_2025_10_21_bj
	31	lmdb_dataset_put_bottles_dustbin_2025_10_22_sh	arrow_dataset_put_bottles_dustbin_2025_10_22_sh
stack_bowls_three	30	lmdb_dataset_stack_bowls_three_2025_10_21_bj	arrow_dataset_stack_bowls_three_2025_10_21_bj
	30	lmdb_dataset_stack_bowls_three_2025_10_21_sh	arrow_dataset_stack_bowls_three_2025_10_21_sh
stack_blocks_three	30	lmdb_dataset_stack_blocks_three_2025_10_21_bj	arrow_dataset_stack_blocks_three_2025_10_21_bj
	31	lmdb_dataset_stack_blocks_three_2025_10_21_sh	arrow_dataset_stack_blocks_three_2025_10_21_sh

1.2 Version 0.2

Task	Episode Num	LMDB Dataset	Arrow Dataset	Visualization
place_shoe	220	lmdb_dataset_place_shoe_2025_09_11	arrow_dataset_place_shoe_2025_09_11
empty_cup_place	196	lmdb_dataset_empty_cup_place_2025_09_09	arrow_dataset_empty_cup_place_2025_09_09
put_bottles_dustbin	199	lmdb_dataset_put_bottles_dustbin_2025_09_11	arrow_dataset_put_bottles_dustbin_2025_09_11
stack_bowls_three	200	lmdb_dataset_stack_bowls_three_2025_09_09 lmdb_dataset_stack_bowls_three_2025_09_10	arrow_dataset_stack_bowls_three_2025_09_09 arrow_dataset_stack_bowls_three_2025_09_10
stack_blocks_three	200	lmdb_dataset_stack_blocks_three_2025_09_10	arrow_dataset_stack_blocks_three_2025_09_10

1.3 Version 0.1

Task	Episode Num	LMDB Dataset	Arrow Dataset
place_shoe	200	lmdb_dataset_place_shoe_2025_08_21 lmdb_dataset_place_shoe_2025_08_27	arrow_dataset_place_shoe_2025_08_21 arrow_dataset_place_shoe_2025_08_27
empty_cup_place	200	lmdb_dataset_empty_cup_place_2025_08_19	arrow_dataset_empty_cup_place_2025_08_19
put_bottles_dustbin	200	lmdb_dataset_put_bottles_dustbin_2025_08_20 lmdb_dataset_put_bottles_dustbin_2025_08_21	arrow_dataset_put_bottles_dustbin_2025_08_20 arrow_dataset_put_bottles_dustbin_2025_08_21
stack_bowls_three	219	lmdb_dataset_stack_bowls_three_2025_08_19 lmdb_dataset_stack_bowls_three_2025_08_20	arrow_dataset_stack_bowls_three_2025_08_19 arrow_dataset_stack_bowls_three_2025_08_20
stack_blocks_three	267	lmdb_dataset_stack_blocks_three_2025_08_26 lmdb_dataset_stack_blocks_three_2025_08_27	arrow_dataset_stack_blocks_three_2025_08_26 arrow_dataset_stack_blocks_three_2025_08_27

2. Usage Example

2.1 LMDB Dataset Usage Example

Ref to RoboTwinLmdbDataset class from robo_orchard_lab. See SEM config for a usage example.

2.2 Arrow Dataset Usage Example

Ref to ROMultiRowDataset class from robo_orchard_lab. Here is some usage example:

2.2.1 Data Parse Example

def build_dataset(config):
    from robo_orchard_lab.dataset.robot.dataset import (
        ROMultiRowDataset,
        ConcatRODataset,
    )
    from robo_orchard_lab.dataset.robotwin.transforms import ArrowDataParse
    from robo_orchard_lab.dataset.robotwin.transforms import EpisodeSamplerConfig
    
    dataset_list = []
    data_parser = ArrowDataParse(
        cam_names=config["cam_names"],
        load_image=True,
        load_depth=True,
        load_extrinsic=True,
        depth_scale=1000,
    )
    joint_sampler = EpisodeSamplerConfig(target_columns=["joints", "actions"])

    for path in config["data_path"]:
        dataset = ROMultiRowDataset(
            dataset_path=path, row_sampler=joint_sampler
        )
        dataset.set_transform(data_parser)
        dataset_list.append(dataset)

    dataset = ConcatRODataset(dataset_list)
    return dataset

config = dict(
    data_path=[
        "data/arrow_dataset_place_shoe_2025_08_21",
        "data/arrow_dataset_place_shoe_2025_08_27",
    ],
    cam_names=["left", "middle", "right"],
)
dataset = build_dataset(config)

# Show all key
frame_index = 0
print(len(dataset))
print(dataset[frame_index].keys())

# Show important key
for key in ['joint_state', 'master_joint_state', 'imgs', 'depths', 'intrinsic', 'T_world2cam']:
    print(f"{key}, shape is {dataset[frame_index][key].shape}")
print(f"Instuction: {dataset[frame_index]['text']}")
print(f"Dataset index: {dataset[frame_index]['dataset_index']}")

# ----Output Demo----
# joint_state, shape is (322, 14)
# master_joint_state, shape is (322, 14)
# imgs, shape is (3, 360, 640, 3)
# depths, shape is (3, 360, 640)
# intrinsic, shape is (3, 4, 4)
# T_world2cam, shape is (3, 4, 4)
# Instuction: Use one arm to grab the shoe from the table and place it on the mat.
# Dataset index: 1

2.2.2 For Training

To integrate this dataset into the training pipeline, you will need to incorporate data transformations. Please follow the approach used in the lmdb_dataset to add the transforms.

from robo_orchard_lab.dataset.robotwin.transforms import ArrowDataParse
from robo_orchard_lab.utils.build import build
from robo_orchard_lab.utils.misc import as_sequence    
from torchvision.transforms import Compose 
train_transforms, val_transforms = build_transforms(config)
train_transforms = [build(x) for x in as_sequence(train_transforms)]

composed_train_transforms = Compose([data_parser] + train_transforms)
train_dataset.set_transform(composed_train_transforms)

2.2.3 Export mcap file and use foxglove to viz

def export_mcap(dataset, episode_index, target_path):
    """Export the specified episode to an MCAP file."""
    from robo_orchard_lab.dataset.experimental.mcap.batch_encoder.camera import (  # noqa: E501
        McapBatchFromBatchCameraDataEncodedConfig,
    )
    from robo_orchard_lab.dataset.experimental.mcap.batch_encoder.joint_state import (  # noqa: E501
        McapBatchFromBatchJointStateConfig,
    )
    from robo_orchard_lab.dataset.experimental.mcap.writer import (
        Dataset2Mcap,
        McapBatchEncoderConfig,
    )

    dataset2mcap_cfg: dict[str, McapBatchEncoderConfig] = {
        "joints": McapBatchFromBatchJointStateConfig(
            target_topic="/observation/robot_state/joints"
        ),
    }
    dataset2mcap_cfg["actions"] = McapBatchFromBatchJointStateConfig(
        target_topic="/action/robot_state/joints"
    )

    for camera_name in config["cam_names"]:
        dataset2mcap_cfg[camera_name] = (
            McapBatchFromBatchCameraDataEncodedConfig(
                calib_topic=f"/observation/cameras/{camera_name}/calib",
                image_topic=f"/observation/cameras/{camera_name}/image",
                tf_topic=f"/observation/cameras/{camera_name}/tf",
            )
        )
        dataset2mcap_cfg[f"{camera_name}_depth"] = (
            McapBatchFromBatchCameraDataEncodedConfig(
                image_topic=f"/observation/cameras/{camera_name}/depth",
            )
        )

    to_mcap = Dataset2Mcap(dataset=dataset)
    to_mcap.save_episode(
        target_path=target_path,
        episode_index=episode_index,
        encoder_cfg=dataset2mcap_cfg,
    )
    print(f"Export episode {episode_index} to {target_path}")


# Export mcap file and use foxglove to viz
dataset_index = dataset[frame_index]["dataset_index"]
episode_index = dataset[frame_index]["episode"].index
export_mcap(
    dataset=dataset.datasets[dataset_index],
    episode_index=episode_index,
    target_path=f"./viz_dataidx_{dataset_index}_episodeidx_{episode_index}.mcap",
)

Then you can use Foxglove and Example Layout to visualize the mcap file. Refer to here to get more visualization example.

3. Data Packer

Data Packaging is the process of parsing the recorded data, performing operations like timestamp alignment, and converting it into a usable format for training.

For mcap data recorded from real-bot, please refer to the released packaging script to perform the conversion.

• Source Code: Packer
• Usage: Instructions