← Leaderboard
VLA-JEPA
VLAUSTC · Feb 10, 2026
VLA pretrained with JEPA-style latent actions learned from video, from USTC with Zhongguancun Academy, SJTU and Tsinghua co-authors.
- Type
- VLA
- Parameters
- —
- Backbone
- Qwen3-VL-2B + V-JEPA2
- Weights
- Open
- License
- Apache-2.0
- Commercial use
- Allowed
- Released
- Feb 10, 2026
- Robots
- Franka Panda (sim), WidowX (Bridge), Google Robot
- SO-101
- –
- LeRobot
- –
Benchmark results
3 of 13 benchmarksLIBERO-PlusSimulation
- Paper / model card ↗Feb 2026
#17 of 2479.5%, self-reported by the authors, not a leaderboard entry
- Paper / model card ↗Feb 2026· n = 2000
#29 of 4697.2%, self-reported by the authors, not a leaderboard entry
- Paper / model card ↗Feb 2026· visual matching
#15 of 1957.3%, self-reported by the authors, not a leaderboard entry
Not reported on 10 benchmarks
Robustness drop: 97.2% on LIBERO → 79.5% on LIBERO-Plus. The perturbed variants show how much of the LIBERO score is memorization.
Adoption
updated Nov 17, 2026Wrong or missing number? Report an error