← Leaderboard

UniVLA (OpenDriveLab)

VLA
OpenDriveLab · May 9, 2025

“Learning to Act Anywhere with Task-centric Latent Actions” (Bu et al., HKU, OpenDriveLab and AgiBot): a VLA trained on task-centric latent actions learned from videos, decoded by a small action head. It is a different model from BAAI’s UniVLA.

Type
VLA
Parameters
—
Backbone
Prismatic-7B
Weights
Open
License
Apache-2.0
Commercial use
Allowed
Released
May 9, 2025
Robots
Franka Panda (sim), WidowX (Bridge), Real-world arm
SO-101
–
LeRobot
–
Inputs
–
VRAM
–

Benchmark results

3 of 13 benchmarks
CALVIN ABC→DNear saturationSimulation
#11 of 113.80 / 5, self-reported by the authors, not a leaderboard entry
LIBEROSaturatedSimulation
#39 of 4695.2%, self-reported by the authors, not a leaderboard entry
SimplerEnv WidowXSaturatedSimulation
#17 of 1947.9%, self-reported by the authors, not a leaderboard entry
Not reported on 10 benchmarks

Adoption

updated Nov 17, 2026
HF downloads · 30d
63
HF downloads · all time
10K
HF likes
10
GitHub stars
1.1K
GitHub forks
71
Wrong or missing number? Report an error