xiaojiahao/libero-10-baseline-subsets

LIBERO-10 baseline-selected subsets

Twelve complete RLDS/TFDS subsets of the modified LIBERO-10 robot demonstration dataset, selected using three baseline objectives at four per-task retention ratios.

Contents

Ratio	Episodes per subset
0.2	80
0.4	157
0.6	231
0.8	308

The source contains 379 demonstrations across 10 tasks. Counts use ceil(ratio * task_demo_count), so the overall fraction is slightly above the nominal ratio.

Directories follow libero_10_no_noops_r{0_2,0_4,0_6,0_8}_{metric}_best. Each directory contains 1.0.0/dataset_info.json, features.json, and complete TFRecord shards, including the original images, observations and actions.

Selection objectives

All three objectives are maximized independently within each task:

• trajectory_quality: normalized covariance-spectrum entropy of action trajectories, truncated to the shortest selected trajectory. This measures action diversity, not smoothness or demonstrated success.
• visual_coverage: normalized covariance-spectrum entropy of flattened sampled visual features.
• vision_action_consistency: signed Pearson correlation between pairwise visual and action cosine similarities. This is not a mutual-information estimator.

Visual objectives use cached OpenVLA features from three uniformly sampled frames per demonstration. The action objective uses full action trajectories; consistency uses actions aligned with the sampled frames.

For each task and ratio, selection chooses the better result from best-pair forward greedy with 10 random-start restarts and random-max search. Random search evaluates up to 200,000 subsets, stopping after 30,000 consecutive attempts without improvement. Greedy additions evaluate all remaining candidates. The seed is task_id * 1009.

“Best” means best found by this heuristic, not a guaranteed global optimum. Ratios are selected independently and are not nested. These subsets have not been evaluated through downstream policy training in this release.

Download and load one subset

Install huggingface_hub, tensorflow, and tensorflow-datasets in your training environment:

from pathlib import Path
from huggingface_hub import snapshot_download
import tensorflow_datasets as tfds

subset = "libero_10_no_noops_r0_4_visual_coverage_best"
local = snapshot_download(
    repo_id="xiaojiahao/libero-10-baseline-subsets",
    repo_type="dataset",
    allow_patterns=[f"{subset}/**"],
)
builder = tfds.builder_from_directory(str(Path(local) / subset / "1.0.0"))
dataset = builder.as_dataset(split="train", shuffle_files=False)

The internal prepared dataset name is liber_o10, matching the source. Use the directory-based loader above for the renamed subset folders. TFRecord data should be read with TensorFlow Datasets; Hugging Face's automatic table preview is not the intended loader.

Verification and provenance

Every source episode was matched to its cached feature metadata using its full action sequence, sampled actions, task description and length. Accelerated scoring was checked against the original baseline functions, including every final winning subset. Exported records passed CRC32C checks and SHA-256 equality checks against the selected source records. Images and episode contents were not modified.

The baseline_selection_reports directory records per-task scores, selected source episode indices, search counts and output sizes. Absolute paths in the reports identify the local generation environment.

Source: OpenVLA modified LIBERO RLDS, whose dataset card specifies the MIT license. Please credit LIBERO and OpenVLA, and refer to those projects for their citations.