docs/data_format.md

Data format

Episode files and indexes

Episodes are grouped by robot, task family, and task configuration:

<robot>/<family>/<case>/
  dataset/
    episode_XXXX.h5
    meta/episode.jsonl
    meta/randomization_episodes.jsonl
  resolved_config.yaml
  results/

• meta/episodes.jsonl lists episodes, unique episode_id values, H5 paths, instructions, and associated files.
• meta/cases.jsonl lists task configurations and their metadata files.
• meta/info.json contains dataset information and current statistics.

Index paths are relative to the dataset root. Episode numbers are local to each task configuration and may be non-consecutive. Use the global index to enumerate files.

For associated annotations, match hdf5_path or meta.episode_index in randomization records and h5_episode_index in result records. Task configurations are available as resolved_config.yaml.

H5 fields

Read array shapes, column names, camera names, and frame rate from each file. Below, T is the frame count, D the selected joint-state dimension, J the full joint-state dimension, K the object count, C the camera count, and S the subgoal count.

Field	Description
agent_pose	(T, D) joint states; columns are described by agent_pose_names and the associated index arrays.
actions	(T, D) joint-position targets.
joint_positions	(T, J) joint states, with columns described by joint_position_names.
ee_pose, ee_pose_valid	End-effector poses and validity flags; see ee_pose_names and ee_frame.
camera_*_rgb	One-dimensional uint8 arrays containing H.264 MP4 video bytes.
language_instruction, subtask_instructions	Episode and subtask instructions.
step_index	(T,) workflow-stage indices; -1 marks transition frames.
object_positions, object_orientations, object_names	Object poses and names. Quaternion order is given by the quaternion_order attribute.
bboxes_2d, bbox_visibility, bbox_camera_names	(T, K, C, 4) image-space boxes, visibility flags, and camera-axis names.
subgoals/frame_idx, subgoals/camera_names, subgoals/images_png	Zero-based frame indices, camera names, and an (S, C) grid of PNG byte arrays.

actions[t] contains the next recorded joint state; the final target is the measured state at episode end. Read the column names and gripper indices when interpreting joint values.

Video bytes must be decoded before accessing images. OpenCV returns BGR images; convert them to RGB before use. Use the H5 fps and frame_count attributes for timing and length.

Read an episode

import json
from pathlib import Path

import h5py

root = Path(".")
with (root / "meta/episodes.jsonl").open() as index:
    item = json.loads(next(index))

with h5py.File(root / item["hdf5_path"], "r") as episode:
    states = episode["agent_pose"][()]
    actions = episode["actions"][()]
    instruction = episode["language_instruction"].asstr()[()]

For an optional LeRobot export with video decoding, see convert_to_lerobot.py and the command in the README.