Episodes Preview Franka Panda Visualizer
20k episodes · 20 fps · 2 cameras · 256×256 av1

LIBERO NVIDIA + LIBERO-plus LeRobot v3 mixture

A LeRobot v3 dataset combining five suites from nvidia/LIBERO_LeRobot_v3 and lerobot/libero_plus. It contains 19,961 episodes, 3,080,750 frames at 20 Hz, and 111 distinct instruction strings. The robot is Franka Panda. The data are intended for later robot policy training; no model weights are included.

Source Pinned revision Episodes Frames
NVIDIA, five suites e5907374380b8f96511957e6ba5582be52a1e179 5,614 842,714
LIBERO-plus f3f49f426d75030177b18778374005bc12ccd588 14,347 2,238,036

Processing

  • The NVIDIA action gripper was converted from 0=open, 1=close to -1=open, +1=close with 1 - 2 * g. LIBERO-plus already uses -1/+1.
  • OpenVLA's episode-aware is_noop() predicate matched zero frames in each source. Thus no frames were removed, and source frame timing and videos were retained.
  • Camera keys were unified to observation.images.front and observation.images.wrist. State has 8 values; action has 7 values. Episode, global frame, and task IDs were remapped to one dataset.
  • Numeric meta/stats.json values, including exact frame-level q01/q99, were recalculated after the gripper conversion. Video statistics were pooled from the source datasets' sampled video statistics. assets/combined/norm_stats.json contains state and action statistics in the OpenPI-style numeric layout.

Simulator replay was not performed. The available LeRobot v3 files do not provide complete per-episode simulator initial states for exact replay. The source trajectories are therefore preserved, but episode success and collision status have not been independently verified. This is recorded as replay_verified: false in meta/build_report.json and every line of meta/source_episodes.jsonl.

Training mixture

The stored frame distribution is not task balanced. meta/mixture.json specifies uniform task selection, then uniform source selection among the sources for that task, uniform episode selection in that source, and uniform valid action-chunk anchor selection in that episode. Use scripts/sample_libero_v3_mixture.py to generate task-balanced anchor indices for horizon 10 or 50. Uniformly sampling rows from the dataset does not apply the task balance.

python scripts/sample_libero_v3_mixture.py \
  --root . --count 111000 --horizon 10 --seed 42 \
  --output anchors-h10.npy

meta/validation_report.json records the full structural validation over all 3,080,750 tabular rows, 19,961 episode intervals, and 108 video files. The output was also spot checked with LeRobot 0.4.4 video and action-chunk loading across the NVIDIA/Plus boundary. The processing and validation scripts are included under scripts/.

Source terms

The NVIDIA source labels its license OpenMDW 1.1. The downloaded lerobot/libero_plus snapshot did not contain a dataset card or license metadata. Check both upstream sources before changing the visibility or redistributing this combined dataset.

Downloads last month
104