§ 01 · Leaderboard
Leaderboard.
Vision-language-action models evaluated on calibrated physical benches. Sorted by ARC composite, continuously re-tested.
ModelArchitecture
01Helix-2Figure · 8.4 B · 2026·02
Hierarchical S2/S1 VLA58.418.247.171.041.6
π0.5Physical Intelligence · 3.3 B · 2025·11
Flow-matching VLA54.116.444.064.538.7
Gemini-Robotics 1.5Google DeepMind · — · 2025·10
Multimodal VLA62.911.241.556.036.4
GR00T N1.5NVIDIA · 12 B · 2025·09
Dual-system VLA51.013.639.860.235.1
RoboBrain-2.0BAAI · 7 B · 2025·08
Embodied-finetuned MLLM + π adapter56.89.636.248.031.9
π0-FastPhysical Intelligence · 3.0 B · 2025·05
Discretized flow VLA47.29.235.752.429.8
Showing 1–6 of 9