ReLAP

VLA from Bagel Labs pretrained on latent actions learned by recurrent video prediction in V-JEPA feature space, then adapted to robot control with a flow-matching action expert. The V-JEPA 2.1 world-model branch is used only for training supervision.

Bagel Labs · Nov 17, 2026 · VLAReport an error
97.9%LIBERO · #41 of 83
72.5%LIBERO+ · #29 of 37
2Benchmarks reported
Nov 17, 2026Released

Benchmark results2

#29 of 3772.5%, self-reported by the authors, not a leaderboard entry
LIBEROSaturatedSimulation
#41 of 8397.9%, self-reported by the authors, not a leaderboard entry
Not reported on 13 benchmarks
Robustness drop: 97.9% on LIBERO → 72.5% on LIBERO-Plus. The perturbed variants show how much of the LIBERO score is memorization.
Wrong or missing number? Report an error