Blame for docs/FORMAT.md

> Repository storage: all episode files are inside TAR shards under `data/`. The episode paths and file schema below describe the layout after extraction. `metadata/episodes.jsonl` adds `archive_path` and `archive_member_prefix`; camera, timestamp and pose semantics are unchanged.

# Data format

The packaged format uses `schema_version: 1.0`. This identifier describes the file schema; the owner-defined dataset release tag is separate. Packaging preserves the exported payloads and normalizes episode metadata as described in CHANGELOG.md.

## Repository layout and identity

`metadata/episodes.jsonl` is the catalog of completed recordings. Each row contains episode_id, task_id, batch_id, calibration_id, episode_path, capture_profile, record_duration_s and export_status. Paths are relative to the repository root. Four task IDs are used: pick_carrot, press_button, open_drawer and hang_cup.

An episode lives in `episodes/<task_id>/<batch_id>/<episode_id>/`. IDs are inherited from the frozen source inventory and are not renumbered by task. Calibration profiles live under `calibration/<calibration_id>/`.

## Episode metadata

`episode.json` specifies source bag name/hash, task and batch IDs, calibration reference, the bag-record origin, all camera-presence flags, video and timestamp-table paths, pose/gripper descriptors, outcome and review labels, encoding options and per-file hashes.

For each camera, `present=false` and `absence_reason=not_recorded` mean the sensor stream was not captured. Video and frame-table paths are null. A recorded stream whose message cannot be decoded remains present and has an explicit failed row; absence and decode failure are different concepts. No placeholder video is created.

The existing per-episode `file_sha256` covers its video, frame-table and Parquet payloads and message definitions. It does not include episode.json itself or shared calibration files. A public release-wide checksum manifest must cover those additional files when packaging is finalized.

## Video frame tables

Each `<camera>.frames.jsonl` has one row per source image message. Important fields:

| Field | Meaning |
| --- | --- |
| source_message_index | Zero-based index within the camera topic in bag-record-time order |
| video_frame_index | Zero-based decoded MP4 display-order index; null on decode failure |
| header_timestamp_ns | Original header timestamp as a decimal integer string; null if absent |
| bag_timestamp_ns | Original recorder timestamp as a decimal integer string |
| record_time_ns | bag_timestamp_ns minus episode_origin_bag_timestamp_ns, as an integer string |
| header_seq, frame_id | Original message fields, when present |
| decode_ok | Whether the image message was decoded and exported |
| video_pts, video_time_base | Encoded playback timestamp and rational time base |
| video_pts_adjustment_ns | Quantization/monotonicity adjustment relative to original record-time differences |
| source_format | Original CompressedImage format string |
| compressed_message_sha256 | Hash of the original compressed image payload |

JSON nanoseconds are strings to avoid precision loss in JavaScript. Do not cast absolute nanoseconds to a floating-point number before subtracting a reference. Source image hashes do not describe the re-encoded MP4 bytes.

## SLAM poses

`slam_pose.parquet` retains every source pose message. Time and index fields use int64; pose and confidence fields use float64. Columns include source_message_index, bag_timestamp_ns, header_timestamp_ns, record_time_ns, header_seq, frame_id, x, y, z, qx, qy, qz, qw, confidence, finite and quaternion_norm.

Positions use metres. Quaternion storage order is xyzw. Quaternions are not silently normalized, and no robot TCP or action conversion is performed. The parent frame_id is retained from the SDK message; the tracked-body-to-camera relationship is explicitly unverified in the release metadata. Confidence values are retained without interpreting them as calibrated probabilities. Non-finite source floating-point values are retained in Parquet and flagged; missing values are distinct from zeros.

## Gripper samples

`gripper.parquet` retains every source sample. Alongside time/index fields, it contains gripper_raw, gripper_open, device_timestamp_raw_s, finite and was_clipped.

`gripper_open = clip(gripper_raw / 88.0, 0.0, 1.0)`. The endpoints 0 and 88 are project-defined normalization references, not precisely calibrated physical limits. Slightly higher readings, such as approximately 89–90 near full opening, are mapped to 1 by design. Clipping alone is not a quality failure or an exclusion criterion. The result is a convenient normalized opening indicator, not millimetres, force or a control action. Raw readings are retained so users can apply another convention. Do not normalize separately using each episode's observed minimum/maximum. Non-finite raw readings are retained, while their derived opening is null.

## Access and dependencies

MP4 uses H.264/yuv420p. Raw grayscale camera channels are recorded in episode metadata; common video decoders may return three-channel arrays. `tools/read_episode.py` is a simple reference reader. Its sequential frame decoder prioritizes unambiguous indexing over optimized random-access throughput. PyArrow reads numeric tables without ROS. Message definitions are included for interpretation, not to require bag parsing.