← Leaderboard

RynnVLA-Latent

VLA
Alibaba DAMO Academy · Oct 25, 2026

Latent-action VLA on DAMO’s RynnBrain VLM, in 2B and 4B sizes. Only the stage-1 latent-action checkpoints are released, which the README says are not closed-loop policies; the 4B model self-reports beating π0.5 and GR00T N1.7 on six real tasks.

Type
VLA
Parameters
5.0B
Backbone
RynnBrain VLM + Mixture-of-Transformers action expert (2B and 4B)
Weights
Open
License
Apache-2.0
Commercial use
Allowed
Released
Oct 25, 2026
Robots
Franka Panda (sim), Aloha-AgileX (sim)
SO-101
–
LeRobot
–
Inputs
–
VRAM
–

Benchmark results

3 of 13 benchmarks
#10 of 2885.6%, self-reported by the authors, not a leaderboard entry
RoboTwin 2.0Near saturationSimulation
  • 88.5%, self-reported by the authors, not a leaderboard entryPaper / model card ↗Oct 2026· self-reported, 50 clean + 500 randomized demos· non-standard
88.5%, self-reported by the authors, not a leaderboard entrynot ranked
LIBEROSaturatedSimulation
#14 of 7098.8%, self-reported by the authors, not a leaderboard entry
Not reported on 10 benchmarks
Robustness drop: 98.8% on LIBERO → 85.6% on LIBERO-Plus. The perturbed variants show how much of the LIBERO score is memorization.
Wrong or missing number? Report an error