← Leaderboard

π0

VLA
Physical Intelligence · Oct 31, 2024

Flow-matching vision-language-action model built on PaliGemma, released with open weights through openpi. It is a common baseline on RoboLab, RoboCasa365, RoboTwin and the LIBERO family.

Type
VLA
Parameters
3.3B
Backbone
PaliGemma 3B
Weights
Open
License
Gemma
Commercial use
Unclear
Released
Oct 31, 2024
Robots
DROID Franka, ALOHA, UR5, SO-100
SO-101
Yes
LeRobot
Supported

Benchmark results

10 of 13 benchmarks
RoboLab-120Simulation
#21 of 225.0%
#25 of 2614.8%
#35 of 401.5%, self-reported by the authors, not a leaderboard entry
#23 of 2456.1%, self-reported by the authors, not a leaderboard entry
#6 of 654.6%, self-reported by the authors, not a leaderboard entry
RoboTwin 2.0Near saturationSimulation
#21 of 2331.4%
LIBEROSaturatedSimulation
#40 of 4694.2%, self-reported by the authors, not a leaderboard entry
SimplerEnv WidowXSaturatedSimulation
#18 of 1927.8%, self-reported by the authors, not a leaderboard entry
RoboArenaReal robot
#12 of 131461
#10 of 121.7%, self-reported by the authors, not a leaderboard entry
Not reported on 3 benchmarks
Robustness drop: 94.2% on LIBERO → 56.1% on LIBERO-Plus → 54.6% on LIBERO-PRO. The perturbed variants show how much of the LIBERO score is memorization.

Adoption

updated Nov 17, 2026
HF downloads · 30d
18K
HF downloads · all time
152.3K
HF likes
32
GitHub stars
14.7K
GitHub forks
2.7K

Running the model

HardwarePrecisionConfigLatency / chunkStepsVRAMChunkControlSource
A100–fine-tuned on ALOHA, 3 images + 14-D state, JAX86 ms––2525 HzFeb 27, 2025 ↗
L40S–RoboLab-120 evaluation (DROID)–––10–May 9, 2026 ↗
RTX 4090–3 cameras, on-board inference73 ms10–5050 HzOct 31, 2024 ↗

Trained on

Runs on

Wrong or missing number? Report an error