rhos-ai/gm100-xtrainer

Dataset Card for GM-100 Xtrainer Part (HDF5 Format)

This is a part of the Great March 100 (GM-100) Project. Raw teleoperation data collected using the Dobot Xtrainer robot. This dataset is stored in HDF5 format.

Dataset Description

This repository contains raw robotic manipulation data stored in HDF5 (.hdf5) format. It is designed for imitation learning tasks using the Dobot Xtrainer hardware stack.

• Robot Platform: Dobot Xtrainer
• Data Format: HDF5
• Camera Views: Top, Left Wrist, Right Wrist

File Structure

The dataset consists of individual HDF5 files, where each file represents one trajectory (episode):

.
├── task_00001/                  # Task 1 Directory
│   ├── train/                   # Training episodes for Task 1
│   │   ├── episode_init_0.hdf5
│   │   ├── episode_init_1.hdf5
│   │   └── ...
│   └── eval/                    # Evaluation/Test episodes for Task 1
│       ├── episode_init_0.hdf5
│       └── ...
├── task_00002/                  # Task 2 Directory
│   ├── ...
└── ...

HDF5 Internal Structure

Each .hdf5 file contains the following groups and datasets. The sequence length is denoted by T.

Root Group

Key	Shape	Dtype	Description
/action	(T, 14)	float	Target joint positions (the command sent to the robot)

Observations Group (/observations)

Key	Shape	Dtype	Description
/observations/qpos	(T, 14)	float	Actual joint positions (state)

Images Group (/observations/images)

Images are stored under /observations/images/.

Note: Images are stored as compressed binary strings (e.g., JPEG buffers) to save space. They must be decoded before use.

Key	Shape	Type	Description
top	(T, 480, 640, 3)	uint8 (binary string)	Top static camera RGB
left_wrist	(T, 480, 640, 3)	uint8 (binary string)	Left wrist camera RGB
right_wrist	(T, 480, 640, 3)	uint8 (binary string)	Right wrist camera RGB

Hardware Specification

Joint Order (14-DOF)

The state and action vectors correspond to the following motors in order:

1. right_waist
2. right_shoulder
3. right_elbow
4. right_forearm_roll
5. right_wrist_angle
6. right_wrist_rotate
7. right_gripper
8. left_waist
9. left_shoulder
10. left_elbow
11. left_forearm_roll
12. left_wrist_angle
13. left_wrist_rotate
14. left_gripper

Usage Example

You can read the data using Python's h5py library.

import h5py
import numpy as np
import cv2
import os

# Example path (adjust as needed)
file_path = "task_00001/train/episode_init_0.hdf5"

if not os.path.exists(file_path):
    print(f"Please check path: {file_path}")
    exit()

with h5py.File(file_path, 'r') as f:
    # 1. Load low-dim data
    qpos = f['/observations/qpos'][:]
    action = f['/action'][:]
    
    print(f"Loaded Episode with {qpos.shape[0]} frames.")
    
    # 2. Load and Decode Images
    # Access the specific camera dataset
    camera_data = f['/observations/images/top']
    
    # Read the first frame's compressed data
    compressed_data = camera_data[0] 
    
    # Decode using OpenCV
    # np.frombuffer converts the binary string to a 1D uint8 array
    image = cv2.imdecode(np.frombuffer(compressed_data, np.uint8), cv2.IMREAD_COLOR)
    
    print(f"Decoded Image Shape: {image.shape}") # Should be (480, 640, 3)

Citation

@misc{wang2026greatmarch100100,
      title={The Great March 100: 100 Detail-oriented Tasks for Evaluating Embodied AI Agents}, 
      author={Ziyu Wang and Chenyuan Liu and Yushun Xiang and Runhao Zhang and Yu Zhang and Qingbo Hao and Hongliang Lu and Houyu Chen and Zhizhong Feng and Kaiyue Zheng and Dehao Ye and Xianchao Zeng and Xinyu Zhou and Boran Wen and Jiaxin Li and Mingyu Zhang and Kecheng Zheng and Qian Zhu and Ran Cheng and Yong-Lu Li},
      year={2026},
      eprint={2601.11421},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2601.11421}, 
}