openroboto-ai/libero-datapool

OpenRoboto Data Pool — LIBERO (v1)

Official training data pool for OpenRoboto (Bittensor subnet 80) miners.

This repository is a byte-for-byte mirror of lerobot/libero at pinned revision a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4, republished unmodified as the canonical, version-pinned training source for subnet submissions.

Contents

• Format: LeRobot v3.0 (parquet + mp4 video streams), robot panda, 10 fps
• 1,693 episodes / 273,465 frames / 40 tasks
• Suites (10 tasks each): libero_spatial, libero_object, libero_goal, libero_10 (long-horizon)

Relationship to the live benchmark

The OpenRoboto benchmark evaluates policies on perturbed variants of these base suites (object / swap / language / task dimensions) with initial states randomized at run time (per-run seed derived from chain block hash + round + drand). Therefore:

• No evaluation BDDL variants, initial-state seeds, or perturbation configs are included here, and none will be published while the current benchmark generation is live.
• LIBERO-90 and LIBERO-Plus are reserved as held-out audit benchmarks and are intentionally excluded from the pool.
• Training on this pool is expected and encouraged — robustness to the eval-time perturbations is what the leaderboard measures.

Revision pinning

Miners must train against the revision advertised in the subnet control.json (dataset.hf_repo + dataset.hf_revision). Published revisions are immutable; pool updates always land as new commits with an explicit revision bump. Scores are comparable only across submissions trained on the same pinned revision.

Quickstart

# requires lerobot >= 0.4 (LeRobot dataset v3 reader)
from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("openroboto-ai/libero-datapool",
                    revision="<pinned revision from control.json>")

License and credits

Apache-2.0, inherited from the source conversion. Original LIBERO benchmark and demonstrations: Liu et al., LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning (2023). LeRobot v3 conversion by the Hugging Face LeRobot team.