openroboto-ai/axis-franka-datapool

OpenRoboto AXIS Franka Datapool

This independently maintained AXIS Franka dataset contains 200 tasks, 35,396 complete episodes and 5,553,077 frames. The expansion on main preserves all original 16,815 episodes, their indices, and their data files. It adds 107 distinct tasks and 18,581 episodes in one new dataset commit. The original release remains available at v0.1-metadata.

Contents and size

This is a metadata-only release: robot states, actions, language instructions, episode metadata, statistics and an exact source selection manifest. It contains no images or videos.

Using the original estimate of 3501.607 AV1 bytes per camera frame and 100 tabular bytes per frame, the selected episodes project to 20.000 GB including future video: 19.445 GB of single-camera AV1 video plus 0.555 GB of tabular data. This is a budget estimate, not the current repository size or a measurement of encoded AXIS video. Any future video addition uses only the 256×256 head camera; wrist/depth streams are excluded from this budget.

Representation

• LeRobot v3.0; nominal source rate 30 Hz, with the source's static-frame filtering retained.
• observation.state: 9D joint position.
• action: 9D joint-position controller target.
• Joint order: [finger_1, finger_2, arm_joint_1, ..., arm_joint_7].
• Original numeric values are retained without normalization, resampling or an action-space conversion.

AXIS and LIBERO have different action spaces, so this dataset is kept separate from openroboto-ai/libero-datapool. loader.py additionally provides the Pi0.5-Axis/DROID 8D adapter at 15 Hz.

Selection and provenance

Source: axisrobotics/Franka-Dataset, pinned to 1d350a3ada19c6201ba2d38a6819583aa04dbcbe. The original tasks use the top-level isaac_state_train Zarr exports. New tasks use the same repository's archives/task_*_isaaclab_state_replay_*.zarr exports, validated for the same joint-position schema and joint order. Archive parts are verified against their published SHA-256 checksums before numeric arrays are extracted.

The baseline selection is retained as an exact prefix. New tasks are selected by a fixed SHA-256 ordering; every selected task receives complete episodes. New-task demonstrations are prioritized, then unused episodes from original tasks fill the remaining frame budget. No episode is duplicated or cut to meet a size target. selection_manifest.json records source paths, task IDs, original episode indices, source frame boundaries, selection seed, the budget calculation and the baseline dataset commit. Its episode order is the release episode order.

OpenRoboto releases this derivative dataset with authorization from the AXIS partner. See TERMS.md for the release notice.

Load

Install lerobot>=0.4, torch, and numpy. Use main for this expansion; pin a dataset commit SHA when exact reproducibility is required.

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    "openroboto-ai/axis-franka-datapool",
    revision="main",
    download_videos=False,
)

With the repository's loader.py:

from loader import load_dataset

dataset = load_dataset(revision="main", action_space="joint_position_target", sample_hz=30)
pi05_dataset = load_dataset(revision="main", action_space="pi05_droid", sample_hz=15)