← Leaderboard

Octo

VLA
UC Berkeley (RAIL) · Dec 2023

Open-source generalist transformer policy from the Octo Model Team (UC Berkeley, Stanford, CMU), pretrained on 800k Open X-Embodiment episodes with a diffusion action head and released in 27M (Small) and 93M (Base) sizes. Weights and code came out in December 2023, and the paper followed in May 2024. Its SimplerEnv and LIBERO numbers here come from evaluations by the SimplerEnv and OpenVLA authors.

Type
VLA
Parameters
93M
Backbone
Transformer trained from scratch + diffusion action head
Weights
Open
License
MIT
Commercial use
Allowed
Released
Dec 2023
Robots
WidowX (Bridge), UR5, Google Robot, Franka, ALOHA (bimanual)
SO-101
–
LeRobot
–
Inputs
RGB · L
VRAM
–

Benchmark results

2 of 13 benchmarks
LIBEROSaturatedSimulation
#67 of 7075.1%, self-reported by the authors, not a leaderboard entry
SimplerEnv WidowXSaturatedSimulation
#27 of 3016.0%, self-reported by the authors, not a leaderboard entry
Not reported on 11 benchmarks
Wrong or missing number? Report an error