← Leaderboard

QuoVLA

VLA
Tsinghua University · May 24, 2026 · built on π0.5

π0.5-style VLA (PaliGemma plus a 300M action expert) using a quotient-space formulation; it reports leading scores on LIBERO and LIBERO-PRO.

Type
VLA
Parameters
2.9B
Backbone
PaliGemma 2.6B
Weights
Unknown
License
–
Commercial use
Unclear
Released
May 24, 2026
Robots
Franka Panda (sim), Aloha-AgileX (sim)
SO-101
–
LeRobot
–

Benchmark results

4 of 13 benchmarks
#4 of 2490.3%, self-reported by the authors, not a leaderboard entry
#1 of 669.8%, self-reported by the authors, not a leaderboard entry
RoboTwin 2.0Near saturationSimulation
  • 51.9%, self-reported by the authors, not a leaderboard entryPaper / model card ↗May 2026· self-reported· non-standard
51.9%, self-reported by the authors, not a leaderboard entrynot ranked
LIBEROSaturatedSimulation
#2 of 4699.6%, self-reported by the authors, not a leaderboard entry
Not reported on 9 benchmarks
Robustness drop: 99.6% on LIBERO → 90.3% on LIBERO-Plus → 69.8% on LIBERO-PRO. The perturbed variants show how much of the LIBERO score is memorization.
Wrong or missing number? Report an error