zelpi.

Physical AI leaderboard

Benchmark
Model size
Show
01 VLA-Adapter-ProAcademic (Wang et al.) · 500M 98.5
02 OpenVLA-OFTStanford · 7B 97.1
03 GR00T N1.6NVIDIA · Size not listed · in zelpi. pool 97.0
04 π0.5Physical Intelligence · Size not listed · in zelpi. pool 96.9
05 UniVLAAcademic (Bu et al.) · 7B 95.2
06 π0Physical Intelligence · 3.3B 94.2
07 GR00T N1NVIDIA · 2B 93.9
08 NORAAcademic (Hung et al.) · 3B 87.9
09 SmolVLAHugging Face · 450M · in zelpi. pool 87.3
10 π0-FASTPhysical Intelligence · 3B 85.5
11 SpatialVLAAcademic (Qu et al.) · 4B 78.1
12 OpenVLAStanford, UC Berkeley & partners · 7B · in zelpi. pool 76.5
13 OctoUC Berkeley & partners · 93M 75.1

Ranked by average LIBERO success. Bars start at 60% so differences are visible. Scores are success rates as reported by the authors or in the cited paper. Evaluation setups differ, so treat close scores as a tie. Sources are linked on the Benchmarks page.