EventVLA

Long-horizon VLA from USTC, Shanghai AI Lab, SJTU and others (CoRL 2026) that keeps an event-driven memory of visual evidence, built on StarVLA’s QwenOFT architecture. The released checkpoints cover RMBench and RoboTwin-MeM, not the 50-task co-train run on the RoboTwin board.

USTC et al. · Jun 18, 2026 · VLA · built on StarVLAReport an error
40.7%RoboTwin 2 · #22 of 33
2.8%Dojo Sim · #51 of 71
2Benchmarks reported
Jun 18, 2026Released

Benchmark results2

#51 of 712.8%
RoboTwin 2.0Near saturationSimulation
#22 of 3340.7%
Not reported on 13 benchmarks
Wrong or missing number? Report an error