disentangled-vla/libero-spatial-island-A-200350-paper200

LIBERO-Spatial Island A (viewpoint 0.20–0.35, paper-200 protocol)

Subset of LIBERO-Spatial demonstrations used as the Island A training half of the shortcut-learning protocol from Xing et al. 2025 (CoRL) "Shortcut Learning in Generalist Robot Policies".

Island A composition

• Tasks: {0, 1, 3, 5, 8} (5 of 10 LIBERO-Spatial tasks)
• Viewpoint normalized range: [0.20, 0.35] (mapped from yaw lower_bound=-10°, upper_bound=90° → angles in [10°, 25°])
• Trajectories: 876 episodes
• Protocol: paper-200, i.e. each task's 200 base demos replayed 4× through the shortcut-renderer with viewpoints sampled uniformly from the island range (demo_repeat_times=4).

The renderer is the paper's modified shortcut-learning-in-grps/LIBERO env, which sets akita_black_bowl_2 to a transparent material (paper §D.2 scene simplification) and overrides the agentview camera yaw per episode.

Format

LeRobot v3 dataset:

meta/info.json
meta/tasks.parquet
meta/episodes/chunk-NNN/file-MMM.parquet
meta/stats.json
data/chunk-NNN/file-MMM.parquet           ← state, action, etc.
videos/observation.images.anchor/...mp4   ← agentview rgb (one mp4 per episode)

• observation.state: 8-D robot proprioception
• action: 7-D end-effector delta + gripper
• observation.images.anchor: 256×256 RGB mp4 (agentview)
• task: free-form language instruction (per episode)

Companion dataset

• Island B half: disentangled-vla/libero-spatial-island-B-650800-paper200

How it was built

scripts/build_island_v3_from_hdf5.py in our repo converts the upstream LIBERO hdf5 demonstrations into LeRobot v3 with per-episode viewpoint sampling drawn from the island range. See repository for the exact command.