VLAct

Qwen3-VL-4B backbone continued-pretrained in the StarVLA codebase on open Franka and AgileX robot data, keeping the vision encoder and lower LLM layers frozen and co-supervising OFT, π and GR00T action heads; downstream tasks attach a fresh action head. Accepted to NeurIPS 2026.

StarVLA Team · Aug 27, 2026 · VLA · built on StarVLAReport an error
83.8%LIBERO+ · #15 of 37
7.6%Dojo Sim · #41 of 71
3Benchmarks reported
Aug 27, 2026Released

Benchmark results3

#41 of 717.6%
#15 of 3783.8%, self-reported by the authors, not a leaderboard entry
RoboTwin 2.0Near saturationSimulation
  • 91.7%, self-reported by the authors, not a leaderboard entryPaper / model card ↗Aug 2026· self-reported, 50 clean + 500 randomized demos· n = 10000· non-standard
91.7%, self-reported by the authors, not a leaderboard entrynot ranked
Not reported on 12 benchmarks
Wrong or missing number? Report an error