← Leaderboard

VLA-Adapter

VLA
OpenHelix Team · Sep 11, 2025

Small VLA that bridges a 0.5B Prismatic VLM to a lightweight policy through a bridge-attention adapter, without robot-data pretraining. A larger-data “Pro” variant reports 98.5 on LIBERO and 4.50 on CALVIN.

Type
VLA
Parameters
500M
Backbone
Qwen2.5-0.5B (Prismatic)
Weights
Open
License
MIT
Commercial use
Allowed
Released
Sep 11, 2025
Robots
Franka Panda (sim), Synria Alicia-D
SO-101
–
LeRobot
–

Benchmark results

2 of 13 benchmarks
CALVIN ABC→DNear saturationSimulation
#9 of 114.42 / 5, self-reported by the authors, not a leaderboard entry
LIBEROSaturatedSimulation
#28 of 4697.3%, self-reported by the authors, not a leaderboard entry
Not reported on 11 benchmarks

Adoption

updated Nov 17, 2026
GitHub stars
2.3K
GitHub forks
215
Wrong or missing number? Report an error