LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation

Paper (arXiv 2610.22372) · Project page · Code · NeurIPS 2026 (Evaluations & Datasets Track)

LIBERO-PeRM measures whether a manipulation policy follows a user preference while completing the task: every episode is scored for task success (SR) and for preference satisfaction (SA) by a per-preference metric at one of three tiers (motion, skill, user profile). Demonstrations are generated by scripted skill controllers in the LIBERO simulator; each mission ships preference-conditioned rollouts and, where the preference can be left out, matched no-preference rollouts.

Release status (2026-10-26)

Suite	Released here	Status
LIBERO-SinglePref	22 tasks, 2200 preference demos + 220 no-preference demos (158 GB)	human-reviewed batch; more tasks are added as they pass review
LIBERO-PrefScale, LIBERO-ComprePref, LIBERO-PrefConflict, LIBERO-PrefCompose	–	upcoming

Each released task was replayed and checked by the authors in the review portal of the code repository (video + satisfaction curve); the demonstrations here are the same batch that was reviewed. manifest.json records sha256, size, episode count and provenance per file.

Tasks in this release

Mission	Task	Preference	Tier / metric	Pref demos	No-pref demos
1-3_107	Put the cream cheese on the bowl	Avoid sudden acceleration	motion / max_acc	100	10
1-7_104	Put the bowl on the plate	Move higher	motion / height_offset	100	10
1-7_114	Pick the chocolate pudding and place it in the basket	Move higher	motion / height_offset	100	10
2-1_019	Put the black bowl on the plate	Pick from right edge	skill / grasp_offset	100	10
2-2_019	Put the black bowl on the plate	Pick from back edge	skill / grasp_offset	100	10
2-2_077	Put the white mug on the left plate	Pick from back edge	skill / grasp_offset	100	10
2-3_077	Put the white mug on the left plate	Pick by handle	skill / grasp_offset	100	10
2-4_119	Pick the salad dressing and place it in the basket	Pick from top at an angle	skill / grasp_tilt	100	10
2-11_118	Pick the orange juice and place it in the basket	Pick from front	skill / grasp_tilt	100	10
3-1_004_1	Put both the alphabet soup and the cream cheese box in the basket	Prefer green objects	profile / plan_choice	100	10
3-1_006_1	Put both the cream cheese box and the butter in the basket	Prefer green objects	profile / plan_choice	100	10
3-10_117_1	Pick up a drink and place it in the basket	Prefer sweet drinks	profile / plan_choice	100	10
3-11_112_1	Pick a sauce and place it in the basket	Prefer sour sauces	profile / plan_choice	100	10
3-12_109_1	Pick a food or drink and put it on the top of the drawer	Prefer boxed objects	profile / plan_choice	100	10
3-15_106_1	Put a bowl on top of the cabinet	Prefer smaller objects	profile / plan_choice	100	10
3-2_001_1	Put the bowl in the bottom drawer of the cabinet and close it	Prefer red objects	profile / plan_choice	100	10
3-4_104_1	Put the bowl on the plate	Prefer blue objects	profile / plan_choice	100	10
3-6_101_1	Put an item inside	Upper drawer for food and drink	profile / plan_choice	100	10
1-2_065	Pick up the alphabet soup and put it in the tray	Slow down	motion / speed_range	100	10
1-2_113	Pick up the butter and put it in the basket	Slow down	motion / speed_range	100	10
1-6_105	Put the bowl on the stove	Move straight	motion / straight_path	100	10
3-7_101_1	Open the one layer of the drawer and put the bowl inside	Middle drawer for kitchenware	profile / plan_choice	100	10

SinglePref/tasksuite.yaml carries the LIBERO instruction, the task text and the preference text per mission; SinglePref/missions/<mission>/ holds the mission configuration (mission_info.yaml: task plan, naive plan, preference metric and its calibration anchors) and the BDDL scene file.

Files

SinglePref/Mission_<mission>.hdf5          preference-conditioned demonstrations (100 per mission)
SinglePref_naive/Mission_<mission>.hdf5    no-preference demonstrations on the same scene (10 per mission)
SinglePref/missions/<mission>/             mission_info.yaml + <scene>.bddl
SinglePref/tasksuite.yaml                  instruction / task / preference text per mission
manifest.json, missions.csv, croissant.json

Drop SinglePref/ and SinglePref_naive/ into VLA_data/ of the code repository and every script there (render_task_gallery.py, calibrate_low_pref.py, the LeRobot converters, policy training) finds them.

HDF5 layout

One file per mission, one group per episode, following the LIBERO demonstration format:

data/demo_<k>/                      attrs: init_state, model_file (MJCF of the sampled scene), num_samples
  actions        (T, 7)   float64   end-effector delta pose (6) + gripper
  dones          (T,)     int64     1 from the step the task predicate holds
  rewards        (T,)     float64
  states         (T, 51)  float64   full MuJoCo state, replayable with sim.set_state
  robot_states   (T, 9)   float64
  obs/agentview_rgb      (T, 256, 256, 3) uint8
  obs/eye_in_hand_rgb    (T, 256, 256, 3) uint8
  obs/frontview_rgb      (T, 256, 256, 3) uint8
  obs/ee_pos, ee_ori, ee_states, gripper_states, joint_states
  low_pref_rewards/<metric>  per-step raw preference signal logged at generation time by the metric
                             version current then; NOT the release metric. Episodes added to a mission
                             later may carry a newer metric version. Recompute preference scores from
                             `states` with the shipped Preference code (see "Preference metrics" below).

import h5py
with h5py.File("SinglePref/Mission_2-1_019.hdf5") as f:
    demo = f["data/demo_0"]
    actions, states = demo["actions"][:], demo["states"][:]
    xml = demo.attrs["model_file"]       # MJCF; see the path note below

Asset paths in model_file. Like the original LIBERO datasets, the stored MJCF references the asset directory of the machine that generated the data. Remap it to your checkout before building a MjSim from the string, e.g. xml.replace("/data3/zhixu/P_Bench/", "<repo>/").replace("/data3/zhixu/LIBERO-PeRM/", "<repo>/") (LIBERO/scripts/calibrate_low_pref.py::_patch_xml in the code repository does exactly this).

Preference metrics

The metric of each mission is named in mission_info.yaml (low_prefs) and implemented in LIBERO/Preference/ of the code repository; render_task_gallery.py replays an episode and plots the satisfaction curve, calibrate_low_pref.py recomputes the preference / no-preference anchors from the demonstrations. Satisfaction is a per-episode score in [0, 100]; the pairing with no-preference rollouts is what makes it discriminative.

License and citation

Data: CC BY 4.0. Code: MIT (see the repository).

@inproceedings{li2026liberoperm,
  title     = {LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation},
  author    = {Li, Zhixu and Tang, Keqian and Gong, Litian and Yao, Jingyu and Zhang, Wenqian and
               Xu, Tianze and Wang, Zehao and Wang, Yuping and Zhang, Junge and Li, Jiachen},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}