saaduddinM/libero_safety_v1

LIBERO Safety

This repository contains 50-scene synthetic LIBERO safety validation sets.

• —v1: original LIBERO safety image set.
• —v2: regenerated set using the latest scene configuration and saved manual positions.
• —v3: same scenes and labels as v2, with a static non-colliding MuJoCo white cutting board fixture added on the table; original V2 object and fixture states are unchanged.
• —v4: corrected zoomed non-reference set with pickable target objects for video rollouts.
• —v5: corrected zoomed white-cutting-board reference set with pickable target objects for video rollouts.

Each config has a train split with rendered scene images and metadata.

Columns:

• —image: rendered LIBERO scene image, inferred from file_name by Hugging Face ImageFolder.
• —scene_id: deterministic scene identifier from scene001 onward.
• —scene_name: scene context name.
• —safe_objects: objects that naturally belong in the scene context.
• —unsafe_object: one object that does not belong in the scene context.
• —video: MP4 pick-place rollout for the same scene. This is present in v4 and v5.
• —reference_object: fixed visual reference object. This is present in v3 and v5.
• —reference_object_category: reference object category. This is present in v3 and v5.
• —reference_object_position: table position for the reference object. This is present in v3 and v5.

Scene counts for each config:

• —giftbox: 17
• —microwave: 16
• —stove: 17

Load with:

from datasets import load_dataset

v1 = load_dataset("saaduddinM/libero_safety_v1", "v1", split="train")
v2 = load_dataset("saaduddinM/libero_safety_v1", "v2", split="train")
v3 = load_dataset("saaduddinM/libero_safety_v1", "v3", split="train")
v4 = load_dataset("saaduddinM/libero_safety_v1", "v4", split="train")
v5 = load_dataset("saaduddinM/libero_safety_v1", "v5", split="train")