Digital-Divide-Data/umi-gripper-demonstration-samples

Sample Data

Each episode is one MCAP bag, three H.264 videos, camera intrinsics, and metadata for scene reconstruction.

Directory layout

episodeNNNN/
├── .__version__                                       format version
├── data/
│   ├── episode.mcap
│   ├── left.perception_interface.camera.state.mp4
│   ├── center.perception_interface.camera.state.mp4
│   └── right.perception_interface.camera.state.mp4
└── meta/
    ├── sensor/
    │   ├── left.metadata_interface.camera_info.meta.yaml
    │   ├── center.metadata_interface.camera_info.meta.yaml
    │   └── right.metadata_interface.camera_info.meta.yaml
    ├── scene/
    │   ├── robot_description.urdf
    │   └── workcell_description.yaml
    └── system/
        ├── mcap.yaml
        ├── video_encoder.yaml
        ├── annotation.yaml
        └── flag.yaml

File	Contents
data/episode.mcap	All 9 topics. No image data.
data/*.mp4	One H.264 video per camera
meta/sensor/*.camera_info.meta.yaml	Intrinsics per camera — see Camera intrinsics
meta/scene/robot_description.urdf	Full kinematic tree: links, inertials, meshes, joint origins, axes, limits, finger mimic relations
meta/scene/workcell_description.yaml	Per-device message type, interface, coordinate frame, declared frequency
meta/system/mcap.yaml	rosbag2 bag info: duration, start time, per-topic message counts
meta/system/video_encoder.yaml	Encoder lib and settings per camera, plus resolution and fps
meta/system/annotation.yaml	annotations: list of {time, annotation}, one task label per episode
meta/system/flag.yaml	flags: list of {time, reason}, or ~ when empty

Topics

Rates as declared in workcell_description.yaml.

Topic	Type	Rate	Contents
/left/perception_interface/pose/state	geometry_msgs/msg/PoseStamped	70 Hz	left pose, 6-DoF
/center/perception_interface/pose/state	geometry_msgs/msg/PoseStamped	70 Hz	center pose, 6-DoF
/right/perception_interface/pose/state	geometry_msgs/msg/PoseStamped	70 Hz	right pose, 6-DoF
/left/perception_interface/gripper/state	sensor_msgs/msg/JointState	200 Hz	one scalar, joint left_gripper
/right/perception_interface/gripper/state	sensor_msgs/msg/JointState	200 Hz	one scalar, joint right_gripper
/left/perception_interface/camera/state	diagnostic_msgs/msg/DiagnosticStatus	60 Hz	frame pointer → left...mp4
/center/perception_interface/camera/state	diagnostic_msgs/msg/DiagnosticStatus	60 Hz	frame pointer → center...mp4
/right/perception_interface/camera/state	diagnostic_msgs/msg/DiagnosticStatus	60 Hz	frame pointer → right...mp4
/tracking/meta_interface/tracking_status/state	std_msgs/msg/String	70 Hz	JSON status payload

Tracking status: The String payload is a JSON document, not a status word. Top level: stale, session_id, sequence, frame_index, sequence_gap, reference_space. Per tracked body — head, left_grip, left_aim, right_grip, right_aim — it carries state, position_valid, orientation_valid, position_tracked, orientation_tracked, location_flags, velocity_flags, linear_velocity, angular_velocity. An input block carries left_index_trigger, left_grip_trigger, right_index_trigger, right_grip_trigger, left_joystick, right_joystick, buttons, touches; buttons and touches are bitmasks, not counts.

Video sync

No image data is in the bag. There is no sensor_msgs/msg/Image message, despite workcell_description.yaml declaring the camera streams as Image. The bag holds one DiagnosticStatus per camera frame instead. Its name field is the mp4 filename, and its values carry:

Key	Meaning
header.stamp	capture time, ns since epoch
video.stamp	capture time, µs since the first frame of that video
height, width	frame resolution, matches video_encoder.yaml
encoding	uyvy (pre-encode pixel format)
is_bigendian	false

To get frame n of a camera: take the n-th DiagnosticStatus on that camera's topic in log-time order and read that index in the mp4.

1. video.stamp is not a frame index. It is a microsecond offset tracking wall time, equal to header.stamp minus the first frame's header.stamp on the same topic.
2. The videos are variable frame rate. Consecutive video.stamp gaps sit near 16.6 ms but are not uniform, so frame_index / 60 drifts against true PTS. The mcap message order is authoritative; read PTS from the container when you need per-frame time.

Flags

flag.yaml holds {time, reason} entries, or ~ when the episode has none. In this dataset the only reason present is frame drops, which is how camera frame drops are logged: each entry marks a moment when one or more cameras failed to deliver a frame at the expected ~16.7 ms cadence, seen as a widened gap in that camera's /…/camera/state topic.

Gripper encoders

The gripper encoders read the revolute joint angle between the angular gripper fingers and return ticks between 0 and 16383 (14 bits resolution). The tick count is normalized to get the gripper angle in radians. (Resolution ~ 4e-4 rads)

Camera intrinsics

meta/sensor/*.camera_info.meta.yaml — one per camera, in sensor_msgs/msg/CameraInfo layout:

Key	Contents
width, height	resolution the calibration was performed at — 1920×1200 for all three, matching video_encoder.yaml
distortion_model	plumb_bob
d	5 distortion coefficients [k1, k2, t1, t2, k3]
k	3×3 intrinsic matrix, row-major flattened
r	3×3 rectification matrix — identity, these are unrectified images
p	3×4 projection matrix, [fx, 0, cx, 0, 0, fy, cy, 0, 0, 0, 1, 0]

p is derived from that camera's own k; fx and cx come from k[0] and k[2], fy and cy from k[4] and k[5].

Extrinsic calibration is separate: it lives in robot_description.urdf as the *_color_optical_frame joints, which are static and valid as written.

TF tree

From robot_description.urdf. world_to_quest_origin is the identity, so quest_origin and world coincide.

world
└── quest_origin                       [fixed]                static
    ├── quest_hmd                      [floating] ← center    dynamic
    │   ├── quest_camera_0             [fixed]                static
    │   ├── quest_camera_1             [fixed]                static
    │   └── center_color_optical_frame [fixed]                static
    ├── quest_left_aim                 [floating] ← left      dynamic
    │   ├── quest_left_aim_visual      [fixed]                static
    │   └── left_base_link             [fixed]                static
    │       ├── left_color_optical_frame [fixed]              static
    │       ├── left_aruco_detect      [fixed]                static
    │       ├── left_left_finger       [revolute] ← encoder   dynamic
    │       └── left_right_finger      [revolute] mimic left  dynamic
    └── quest_right_aim                [floating] ← right     dynamic
        ├── quest_right_aim_visual     [fixed]                static
        └── right_base_link            [fixed]                static
            ├── right_color_optical_frame [fixed]              static
            ├── right_aruco_detect     [fixed]                static
            ├── right_left_finger      [revolute] ← encoder   dynamic
            └── right_right_finger     [revolute] mimic right dynamic

Three floating joints carry the only 6-DoF transforms, one per pose topic. Four revolute finger joints come from the two encoder topics; each gripper logs one scalar and the second finger is a URDF mimic of the first, not measured separately. The *_aruco_detect frames are static references with no data attached.