LIBERO-X Levels 1-3, 218-task subset (LeRobot v3.0)
A subset of the human teleoperation demonstrations of LIBERO-X, converted to the LeRobot v3.0 dataset format. It contains the demonstrations whose instruction matches one of the 218 Level 1-3 tasks (152 distinct instructions) used in our LIBERO-X experiments.
Papers
- Failure Prediction and Classification in Vision-Language-Action Models Via Calibrated Latent Probing
- LIBERO-X, the source of the demonstrations (RSS 2026)
| Episodes | 1,076 (2 to 44 per instruction) |
| Frames | 371,555 at 10 fps |
| Instructions | 152 distinct (the 218 tasks share instructions across scenes) |
| Cameras | image (agent view) and wrist_image, 256x256 RGB, stored in the parquet files |
| State / action | 8 / 7 float32 (see "Conventions") |
| Format | LeRobot v3.0, 506 parquet data files, 39 GB |
| Robot | Franka Panda, simulation (robosuite / LIBERO) |
Provenance, license and changes
The data comes from LIBERO-X by Meituan (dataset, code, paper, accepted at RSS 2026), released under CC BY 4.0. This subset is released under the same license. If you use it, please cite the LIBERO-X paper (BibTeX: see the paper page) and credit the original authors. This subset is not endorsed by them.
Changes made to the original (the images, states and actions themselves are unchanged):
- Selection: only episodes whose instruction text (whitespace-stripped) equals one of the 152 instructions of the 218 tasks (see "Why these tasks"). The original has 2,520 episodes.
- Format: converted from LeRobot v2.1 to v3.0 with LeRobot's own converter (0.7.0). Episodes are packed into larger parquet files.
- Renumbering: episodes are renumbered 0 to 1,075 and the global
indexis recomputed. The map to the original numbering is insource/episode_map.json. - Tasks: the original has 41 instructions that appear twice under different
task_indexvalues (and one with a trailing space). They are merged by text, so there are 152 tasks, renumbered 0 to 151. The originaltask_indexof each episode is kept insource/episode_map.json.
Why these tasks
The 218 tasks are the finetune_set_v2 selection: LIBERO-X extension tasks (EXTENSION_KITCHEN / EXTENSION_STUDY
scenes) of levels 1-3 that are backed by human demonstrations (72 Level 1, 73 Level 2, 73 Level 3). Each task has a
skill (8 in total: place_on_surface, place_in_container, place_relative_to_landmark, stack_objects,
articulate_open, articulate_close, toggle_appliance_on, toggle_appliance_off) and a signature
(skill | target object | spatial relation | landmark | single or compound instruction; 72 in total). To keep the set
balanced, the two dominant placement skills are capped at 3 task instances per signature, while the under-represented
skills are uncapped (every demo-backed instance is included), and the three levels are roughly even. The set is
kitchen-heavy (194 kitchen and 24 study tasks).
The full per-task table is splits/splits.csv (level, skill, sig, room, scene, target_class, relation, landmark, compound, n_demos, file, lang and the split columns). n_demos there was counted by instruction text alone, so
tasks that share an instruction share the same demos, and the column over-counts if summed.
Splits
In our experiments the VLA policy was fine-tuned on all tasks, and three seen/unseen splits over the 218 tasks were used to test the generalization of a latent probe. In principle the same splits can be used to test the generalization of a policy itself. They are in increasing difficulty:
| split | what is unseen | what has been seen |
|---|---|---|
leave_task_out (175 seen / 43 unseen tasks) |
20 % of the task instances (scene + instruction), stratified by skill (seed 0) | the same skill and, for most unseen tasks, the same signature |
leave_signature_out (162 / 56) |
16 whole signatures (about 20 % per skill, at most half of a skill's signatures) | the skill, but not that object / relation / landmark combination |
leave_skill_out (8 folds) |
one of the 8 skills entirely | the other skills only; the primitive itself is new |
On leave_signature_out, one unseen signature (stack_objects | bowl | on | bowl | single) holds 16 of the 17
stack_objects tasks, so that skill has a single seen task.
The splits are defined on tasks (scene + instruction), but the episodes only carry the instruction text. Episodes are
therefore assigned to splits through their instruction, in splits/episode_splits.csv (one row per episode) and
splits/instruction_splits.csv (one row per instruction):
leave_signature_outandleave_skill_outare clean: every instruction belongs to exactly one signature and one skill, so each episode is cleanly seen or unseen (train/test).leave_task_outis not clean at the episode level: 19 of the 152 instructions are used by both a seen and an unseen task (in different scenes). Their episodes are labelledmixed. Only the 18 test-only instructions are guaranteed unseen. Decide how to treatmixedepisodes (for example drop them).
Contents
data/chunk-*/file-*.parquet frames (images, state, actions, indices)
meta/ LeRobot v3.0 metadata (info.json, tasks.parquet, stats.json, episodes/)
splits/splits.csv the 218 tasks with skill, signature, level, scene and split labels
splits/instruction_splits.csv one row per instruction (152): skill, signature, levels, split labels, episode count
splits/episode_splits.csv one row per episode: its instruction and split labels
source/episode_map.json new episode index -> original LIBERO-X episode and task index
Conventions
state(8): end-effector positionx, y, zin metres, end-effector orientation as an axis-angle vector (the original metadata calls it roll/pitch/yaw), and the two gripper finger positions.actions(7):[dx, dy, dz, drot_x, drot_y, drot_z, gripper], controller commands for robosuite'sOSC_POSE(delta end-effector pose, normalised to about [-1, 1]);gripperis -1 (open) or +1 (close).- The world frame is centred on the table (not on the robot base) and is the same for all scenes: +x points from the robot across the table. The base is at x of about -0.66 m (kitchen) or -0.75 m (study).
- Actions are teleoperation commands, not differences of the states.
Known quirks
- A few action outliers: 0.74 % of frames have an action component above 1 in magnitude, and the extremes reach -20 and +9 (in 10 episodes above 5). Consider clipping or filtering, especially with mean/std normalisation.
Loading
from lerobot.datasets import LeRobotDataset
ds = LeRobotDataset("podolinsky/libero-x-level123") # needs access to this repo
print(ds.num_episodes, ds.num_frames) # 1076 371555
Related: the rollouts, embeddings and failure labels of a GR00T policy on these 218 tasks are in
podolinsky/gr00t-libero-x-features-v3.
Note: demonstrations are labelled by instruction text only in LIBERO-X, not by scene, so episode-level labels for leave_task_out are approximate.
- Downloads last month
- 156
