Benchmarks/LIBERO-Plus

LIBERO-Plus

SimulationCode checkReleased Oct 2025Index weight 0.595
LIBERO-Plus authorsMaintainer ↗

Seven perturbations, camera, robot start, language, lighting, background, sensor noise and layout.

As of November 17, 2026, OpenVLA-OFT+ leads with 79.6, followed by OpenVLA-OFT (69.6) and RIPT-VLA (68.4). Official leaderboard ↗ · Paper ↗

Leaderboard

Tasks
10,030
Attempts / task
1
Median cost / task
$0.011
Median time / task
12 s

Success rate (%) under each perturbation.

#Model
1OpenVLA-OFT+Not counted in the index79.692.830.385.8——
2OpenVLA-OFT69.656.431.979.5$0.0098est.11 sest.3RIPT-VLA68.455.231.277.6$0.0099est.11 sest.
4OpenVLA-OFT_mNot counted in the index67.955.621.781.0——
5π0-FAST61.665.121.661.0$0.010est.11 sest.
6OpenVLA-OFT_wNot counted in the index55.810.438.770.5——
7π053.613.86.058.8$0.011est.12 sest.8UniVLA43.91.846.269.6$0.012est.13 sest.9NORA39.02.237.065.1$0.012est.13 sest.10WorldVLA25.00.1027.941.6$0.013est.14 sest.11OpenVLA15.60.803.523.0$0.014est.15 sest.

Tasks

10,030 perturbed tasks; one per suite and perturbation shown. Hover a task to play its clip (28 of 28).

28 tasks · 28 with clips

Methodology

Scoring
Code check
Time cap / attempt
17 s
Control rate
20 Hz
Models in index
8 of 11

Scoring. BDDL goal predicates checked in the simulator, pass/fail.

Who runs it. Benchmark authors.

Robots. Simulated Franka.

Time per attempt. Not published. Estimated otherwise: time cap 16 s × (1 − success/2).

Cost per attempt. Not published. Estimated otherwise: one H100 at $3.29 per hour (Lambda on-demand, Oct 2026) × time. GPU price source ↗