README.md

metadata

license: other
task_categories:
  - robotics
tags:
  - LeRobot
  - franka
  - robot-learning
  - joint-space
pretty_name: OpenRoboto AXIS Franka Datapool

OpenRoboto AXIS Franka Datapool

This is OpenRoboto's independently maintained AXIS Franka release. It is kept separate from openroboto-ai/libero-datapool because AXIS and LIBERO do not share an action space.

The v0.1-metadata release contains observations, actions, language tasks, and episode metadata only. It deliberately contains no rendered or encoded video. selection_manifest.json fixes the exact source episodes, so a later single-camera AV1 addition can use the same split. Its projected total size, including that future video, is approximately 10 GB.

The fixed split contains 16,815 episodes and 2,776,873 frames from 93 AXIS V1 tasks. The projection is 9.723 GB of future AV1 video plus 0.278 GB of tabular data, or 10.001 GB total.

Representation

• Format: LeRobot v3.0, 30 Hz
• observation.state: 9D joint position
• action: 9D joint-position controller target
• Layout: [finger_1, finger_2, arm_joint_1, ..., arm_joint_7]
• Camera in the source: one 256×256 head_camera (not included in this release)

The source representation is canonical and lossless. We do not pre-convert it to LIBERO's 7D end-effector action, because forward kinematics cannot recover the environment-specific OSC command exactly. loader.py instead offers the 8D Pi0.5-Axis/DROID adapter used by the AXIS training recipe.

Load

from loader import load_dataset

# Exact 9D source representation at 30 Hz
dataset = load_dataset(action_space="joint_position_target", sample_hz=30)

# Official-style online adapter: 8D DROID view, sampled at 15 Hz
pi05_dataset = load_dataset(action_space="pi05_droid", sample_hz=15)
sample = pi05_dataset[0]

Install lerobot>=0.4, torch, and numpy. Standard LeRobot loading also works directly:

from lerobot.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset(
    "openroboto-ai/axis-franka-datapool",
    revision="v0.1-metadata",
    download_videos=False,
)

Provenance and reproducibility

Source: axisrobotics/Franka-Dataset, pinned to the commit recorded in meta/info.json and selection_manifest.json. Selection is deterministic, episode-level, and records every original (task_id, episode_index) pair. OpenRoboto is releasing this derivative dataset with authorization from the AXIS partner. See TERMS.md for the release notice; no third-party personal account is used to publish this repository.