Labelbox/robotics-datasets

LBX Robotics Datasets

A preview of Labelbox robotics data offerings.

Bimanual xArm Manipulation

Bimanual teleoperated manipulation recordings collected on two dual-arm xArm stations. Sessions cover a broad set of household and table-top tasks and are shared as raw MCAP files so the full sensor and control fidelity is preserved end-to-end.

Robots

Two bimanual stations are included:

Station	Arms	URDF
Bimanual xArm6	2× UFACTORY xArm6 (6 DoF each) with parallel grippers	urdf/dual_xarm6.urdf
Bimanual xArm7	2× UFACTORY xArm7 (7 DoF each) with parallel grippers	urdf/dual_xarm7.urdf

Each station carries three ZED 2i cameras: one on the left wrist, one on the right wrist, and one overhead looking down at the workspace (the overhead camera's left and right stereo views are exposed as separate topics). Visual STL meshes for both arms, the parallel gripper, the end-tool plate, and the ZED 2i mount are provided so you can render the robot in a 3D viewer.

Data format

All recordings are MCAP files with protobuf-encoded messages. They open directly in Foxglove Studio for visual inspection, or via the mcap Python package. Cameras are recorded at roughly 30 fps; robot proprioception and control signals at higher rates (see Topics below).

Folder layout

Labelbox/robotics-datasets
├── README.md
└── data/
    └── bimanual/
        ├── bimanual_foxglove_layout.json     # Foxglove layout for these recordings
        ├── urdf/                             # robot models + STL meshes
        │   ├── dual_xarm6.urdf
        │   ├── dual_xarm7.urdf
        │   └── meshes/
        │       ├── xarm6/visual/*.stl
        │       ├── xarm7/visual/*.stl
        │       ├── gripper/xarm/*.stl
        │       ├── end_tool/collision/*.stl
        │       ├── zed2i.stl
        │       └── zed_camera_mount_xarm.stl
        ├── xarm6/
        │   └── <task_name>/
        │       └── recording_NNNN.mcap
        └── xarm7/
            └── <task_name>/
                └── recording_NNNN.mcap

Tasks

xArm6 station Task Files fold_tshirt 13 scrub_plate_with_sponge 24 place_clothing_into_basket 27 place_pot_on_stove_and_add_food_to_pot 2 remove_towel_from_drawer 2 watering_house_plant 2 put_trash_into_trash_bin 2 remove_hangers_from_rack 2 remove_hats_and_dirty_laundry 2

xArm7 station Task Files remove_lunch_box_from_backpack 30 separate_condiments_baby_food_and_spices 12 set_up_dinner_table 27 place_objects_into_drawer 7 place_batteries_into_container 2 place_toothbrushes_into_holder 2

Topics

Topics present in every recording, with approximate publish rates from a representative session (rates vary slightly across recordings).

Topic	Approx. rate	Description
/left-arm-proprio, /right-arm-proprio	~245 Hz	Joint state of each arm
/left-eef-proprio, /right-eef-proprio	~245 Hz	End-effector state (pose, gripper)
/left-arm-leader, /right-arm-leader	~73 Hz	VR controller state from the teleop leader
/left-arm-ik-solution, /right-arm-ik-solution	varies with motion	Inverse-kinematics solver output
/left-arm-tf, /right-arm-tf	~19 Hz	Per-arm transforms
/left-wrist-camera/image-raw	~30 Hz	Left wrist ZED 2i
/right-wrist-camera/image-raw	~30 Hz	Right wrist ZED 2i
/top-left-camera/image-raw	~30 Hz	Overhead ZED 2i, left stereo view
/top-right-camera/image-raw	~30 Hz	Overhead ZED 2i, right stereo view
/{...}-camera/camera-info	once per file	Camera calibration
/tf-static	once per file	Static transforms
/subtask-annotation	sparse	Semantic step labels within a session

All messages use protobuf encoding. Robot-specific schemas are namespaced under lbx_robotics.msg.* (e.g. ArmState, EndEffectorState, IKSolution, VRControllerState, SubtaskAnnotation). Camera and transform schemas use the standard foxglove.* types.

Embedded session metadata

Every MCAP file contains a session-metadata record with the following fields:

Field	Description
task-id	Canonical task name (matches the folder name, e.g. fold_tshirt)
task-instruction	Human-readable instruction shown to the operator
operator-id	Opaque UUID identifying the teleop operator (no PII)
station-id	Identifier of the physical station the data was collected on
session-uuid	Unique session identifier
start-time-unix, end-time-unix	Session start and end times (seconds since epoch)
urdf	The full URDF used during the session, embedded inline as XML

Viewing in Foxglove

A Foxglove layout is provided at data/bimanual/bimanual_foxglove_layout.json. It pre-configures a 3D panel with the robot URDF, four image panels for the cameras, and raw-message panels for the proprioception and teleop streams.

1. Download data/bimanual/urdf/ and data/bimanual/bimanual_foxglove_layout.json to your machine.
2. Open any .mcap recording in Foxglove Studio.
3. View → Import layout from file… and select bimanual_foxglove_layout.json.
4. In the 3D panel, find the URDF layer and set its filePath to your local copy of dual_xarm6.urdf (for xArm6 recordings) or dual_xarm7.urdf (for xArm7 recordings). The layout ships with this field intentionally blank.

License

Released under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license. You are free to share and adapt the material for any purpose, including commercially, with appropriate credit.

Suggested attribution:

Labelbox Robotics. LBX Robotics Datasets — Bimanual xArm Manipulation. Hugging Face, 2026. https://huggingface.co/datasets/Labelbox/robotics-datasets

Contact

Maintained by the Labelbox Robotics team. For questions, collaborations, or feedback, use the dataset's Discussions tab on Hugging Face.