franka-demos

Franka Emika teleoperated demonstrations recorded as ROS 2 bags, with a synchronised HDF5 conversion.

RGB is stored decoded as image arrays (lzf-compressed) — ready to index without a JPEG decode step. A smaller variant that keeps RGB as JPEG bytes lives at fm-dev/franka-demos-compressed (~4x smaller).

Contents

raw/<task>/<operator>/<episode>/   # original ROS 2 bags (.db3.zstd) + metadata.yaml + task.txt
h5/<task>/<episode>.h5             # converted, time-synchronised HDF5
annotations/<version>/...          # machine-generated labels and their QA reviews

Task	Episodes	Converted HDF5	Prompt
button_order	51	51	press the buttons with the same order shown in the video
pick3	50	50	pick 3 times
shuffle	57	0	shuffle the cup with the cube, then press the button with the cube
total	158	101

Decoded HDF5 coverage is partial. The decoded variant was produced for the earlier tasks only; shuffle ships rosbags and annotations here but its decoded HDF5 was not published, because the decoded encoding is ~4x larger and fully reproducible from the rosbags in this repository. The complete converted set for every task is available as JPEG-RGB at fm-dev/franka-demos-compressed.

Rebuild the decoded variant locally with:

python3.10 convert_all.py --mode decoded   # see fm-dev/docs

Episodes rejected by quality control are not published: recordings with fully NaN proprioception, incomplete task executions, and one unrecoverable truncated archive were withheld, so the counts above are the retained set rather than everything that was recorded.

H5 layout

One group per synchronised sample, frame_000000 ... frame_NNNNNN, each containing:

• Cameras (cam_base, cam_hand): colour image, aligned_depth_to_color as uint16 (720x1280), and the matching camera_info.
• Franka state: current_pose (position / orientation), measured_joint_states, desired_joint_states, external_joint_torques, external_wrench_in_base_frame, external_wrench_in_stiffness_frame, desired_end_effector_twist, last_desired_pose.
• Commands: _target_pose_position, _target_pose_orientation, _target_joint.
• Gripper / teleop: _franka_gripper_joint_states, _spacenav_joy_buttons.
• Timing: original_timestamps, camera_timestamps, other_timestamps, plus a sync_quality group with the per-topic offset from the reference RGB stream, a stale_topics list and a depth_stale flag.

Frames are synchronised onto the cam_base colour stream at full rate (no decimation). Non-reference topics are matched within +-100 ms and otherwise best-effort filled from the nearest message (capped at 1 s) with the frame flagged stale, so a good RGB frame is never dropped because one stream skipped a beat. Sample 0 of each episode is dropped.

Converted with rosbag_to_h5_simple.py.

import h5py
with h5py.File("h5/pick3/pick3_20260826_205000_123.h5", "r") as f:
    g = f["frame_000000"]
    rgb = g["_cam_base_camera_color_image_raw_compressed"][:]   # HxWx3 array
    depth = g["_cam_base_camera_aligned_depth_to_color_image_raw_compressedDepth"][:]

Annotations

annotations/<version>/... holds machine-generated episodic annotations: Writer keyframe/subgoal sidecars, offline semantic-event and phase labels, and the QA reviews that assessed them. Each version keeps its own model, prompt version and review verdict; versions are never merged and a superseded run is kept as-is rather than deleted.

Version	Files
events-agent-reviewed-v1	1
events-astra-refinements-v3	42
events-astra-v1	35
events-astra-v2	1470
events-candidates-v1	160
phases-agent-reviewed-v1	109
protocols	1
qa	473
writer-astra-pick3-pilot-v1	2
writer-astra-real-v1	111
writer-astra-real-v2	52
writer-astra-repairs-v1	3
writer-astra-shuffle-pilot-v1	2
writer-luna-native-efficiency-v3	2
writer-luna-workspace-v4	2
writer-terra-current-cli-v5	2
writer-terra-standard-v6	2

annotations/EXPORT_MANIFEST.json records, per bundle, the producer's own complete, human_ground_truth and training_ready flags, plus what the export deliberately omits.

None of these labels is human ground truth, and no bundle is certified training-ready. Some versions were explicitly rejected by review and are published only so the comparison stays auditable. Read the QA verdicts under annotations/qa/ before using any of it for supervision. Per-decision generation traces (storyboards and request/response logs) are not published.