LIBERO-Plus

Last updated: Oct 7, 2026

On this page

Dimensions Data composition Metrics Leaderboard How to run

LIBERO-Plus (CVPR 2026, Fudan / FudanNLP) is a test-only robustness overlay on the original LIBERO Spatial / Object / Goal / Long tasks, not a fifth lifelong suite. The authors generate 10,030 instances with controlled perturbations on seven factors (21 sub-dimensions):

Object layout — target displacement, confounders (1,525).
Camera viewpoint — pose and FOV (1,599).
Robot initial state — start pose (1,550).
Language instruction — LLM rewrites (1,537).
Lighting — intensity, direction, color, shadow (1,142).
Background texture — scene / surface look (1,076).
Sensor noise — photometric degradation (1,601).

Instances are binned into L1–L5 by how many of four reference models solved them (L1 = all four, L5 = none). The paper's scan of ten VLAs reports drops from about 95% on clean LIBERO to below 30% under modest camera or initial-state shifts — an upstream finding, not a hub score.

Paper: LIBERO-Plus. Project: sylvestf.github.io/LIBERO-plus. Code: sylvestf/LIBERO-plus (MIT; drop-in libero fork).

WorldFoundry is official-results only. You run the upstream fork, then official-validation normalizes the export. No in-tree robosuite, no Plus assets. Base LIBERO can wire closed-loop execution; this card cannot.

Dimensions

Object layout
KindAxis
Target displacement and confounding objects (1,525 instances).
Camera viewpoint
KindAxis
Camera position, orientation, and field of view (1,599 instances).
Robot initial state
KindAxis
Manipulator start-pose variation (1,550 instances).
Language instruction
KindAxis
LLM-rewritten task wording (1,537 instances). Still a visual/physical perturbation suite, not LIBERO-Para.
Lighting
KindAxis
Intensity, direction, color, and shadow (1,142 instances).
Background texture
KindAxis
Scene and surface appearance (1,076 instances).
Sensor noise
KindAxis
Photometric distortion and image degradation (1,601 instances).
Difficulty L1–L5
KindTrack
Empirical bins from four reference models: L1 solved by all four, L5 by none.

Data composition

Size
10,030 test-only perturbed instances over the original LIBERO Spatial / Object / Goal / Long tasks. Seven factors, 21 sub-dimensions, five difficulty levels. Official assets: sylvestf/LIBERO-plus (drop-in libero fork).
Sources
github.com
Splits
spatial · object · goal · long · l1 · l2 · l3 · l4 · l5
Modalities
language · vision · action · state
License
mit
Notes
Paper analysis of ten models reports drops from ~95% to below 30% under modest camera / initial-state shifts. That is an upstream finding, not a WorldFoundry board.

Metrics

Metric

success_rate
DirectionHigher is better
Episode success on the perturbed instance, imported from the official LIBERO-Plus evaluator.
PrimarySuccess Rate
task_success
DirectionHigher is better
Per-task or per-factor success among Spatial / Object / Goal / Long cells.
episode_success
DirectionHigher is better
Per-episode success indicator when the export lists individual rollouts.
action_accuracy
DirectionHigher is better
Offline action-prediction accuracy if the official dump includes it.

Leaderboard

Results from the paper (not a live board; not WorldFoundry-run scores). LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models (2025).

Model	Total success (%)
OpenVLA-OFT+paper “Ours”	79.5
OpenVLA-OFT	69.6
RIPT-VLA	68.4
OpenVLA-OFT_m	67.9
π0-Fast	61.6
π0	53.6
NORA	39.0
OpenVLA	15.6

How to run

worldfoundry-eval zoo benchmark-run \
  --benchmark-id libero-plus \
  --mode official-validation \
  --official-results-path <official_results.json> \
  --output-dir tmp/libero-plus/official-validation \
  --json

Optional WORLDFOUNDRY_LIBERO_PLUS_ROOT points at the sylvestf/LIBERO-plus checkout. Default import path: WORLDFOUNDRY_LIBERO_PLUS_RESULTS_PATH. Manifest: worldfoundry/data/benchmarks/tasks/external/libero-plus.yaml. Outputs: scorecard.json (primary success_rate), raw_results.jsonl, and evaluation/summary.json.

← Benchmark Hub