zhourui9813/DexRoam_Realworld_Data

DexRoam Push Chair and Close Laptop Real-World Data

This repository contains real-robot demonstrations of an Astribot robot pushing a chair and then closing a laptop. The data was collected for DexRoam and is provided as a working example for validating raw-data loading and VLA training pipelines.

This release is an integration example rather than a benchmark or a complete training corpus. Results trained only on this subset should not be interpreted as representative of the full DexRoam system.

Dataset summary

Item	Value
Robot	Astribot mobile manipulation platform
End effectors	Dual XHand 12-DoF dexterous hands
Task	Push the chair and close the laptop.
Source format	One HDF5 file per episode
Episodes	50
Frames	37,422
Control/capture frequency	Approximately 30 Hz
Total recorded duration	Approximately 20.7 minutes
Episode duration	21.6–33.7 seconds
Raw size	Approximately 227.3 GiB
Cameras used for training	Head-mounted stereo RGB cameras

The released source episode IDs are contiguous and range from 0 to 49.

Directory structure

DexRoam_Realworld_Data/
├── README.md
└── Astribot-Xhand-data/
    └── robot_push_chair_and_close_laptop/
        ├── push_chair_and_close_laptop_episode_0.hdf5
        ├── ...
        └── push_chair_and_close_laptop_episode_49.hdf5

Raw HDF5 format

Every dataset in an episode has the same leading time dimension T. Episode length varies, but the field names, dtypes, and non-time dimensions are consistent across the release.

The main fields are:

HDF5 key	Shape	Description
/time	(T,)	Monotonic frame timestamps
/images_dict/stereo_left/rgb	(T, 480, 640, 3)	Left stereo RGB stream
/images_dict/stereo_right/rgb	(T, 480, 640, 3)	Right stereo RGB stream
/images_dict/head_stereo/rgb	(T, 480, 1280, 3)	Concatenated stereo stream
/images_dict/head/rgb	(T, 720, 1280, 3)	Additional monocular head RGB stream
/joints_dict/joints_position_state	(T, 25)	Robot joint position state
/joints_dict/joints_velocity_state	(T, 25)	Robot joint velocity state
/joints_dict/joints_position_command	(T, 25)	Robot joint position command
/dex_hand_dict/{left,right}/state	(T, 12)	Left/right dexterous-hand state
/fk_poses_dict_from_state/{astribot_arm_left,astribot_arm_right,astribot_torso}	(T, 7)	FK pose derived from robot state
/fk_poses_relative_action/{astribot_arm_left,astribot_arm_right,astribot_torso}	(T, 7)	Relative Cartesian action
/mobile_dict/absolute_position	(T, 3)	Mobile-base absolute pose
/mobile_dict/relative_position	(T, 3)	Mobile-base relative action
/human_hand_dict/{left,right}/raw_data	(T, 175)	Auxiliary raw hand-tracking data

Pose vectors use [x, y, z, qx, qy, qz, qw]. The provided converter normalizes the quaternion and maps it to the 6D rotation representation used by the training code. It also converts the XHand joint angles from degrees to radians.

The RGB arrays follow the Astribot SDK convention and are stored as BGR byte arrays in the HDF5 files. Use the provided converter to avoid channel-order mismatches.

Inspect one episode

import h5py

path = (
    "Astribot-Xhand-data/robot_push_chair_and_close_laptop/"
    "push_chair_and_close_laptop_episode_0.hdf5"
)

with h5py.File(path, "r") as episode:
    print(episode["time"].shape)
    print(episode["images_dict/stereo_left/rgb"].shape)
    print(episode["joints_dict/joints_position_state"].shape)