← Leaderboard

CogACT

VLA
Microsoft Research Asia · Nov 29, 2024 · built on OpenVLA

VLA from Microsoft Research Asia and Tsinghua University that conditions a diffusion-transformer action module on the output of a 7B VLM initialized from OpenVLA, and smooths its actions with adaptive action ensembling. Its SimplerEnv numbers are self-reported, as are results on the authors’ Realman and Franka setups.

Type
VLA
Parameters
—
Backbone
Prismatic 7B VLM (Llama 2, DINOv2 + SigLIP) + DiT action module
Weights
Open
License
MIT
Commercial use
Allowed
Released
Nov 29, 2024
Robots
Google Robot (sim), WidowX (sim), Realman arm, Franka
SO-101
–
LeRobot
–
Inputs
RGB · L

Benchmark results

1 of 13 benchmarks
SimplerEnv WidowXSaturatedSimulation
#21 of 3051.3%, self-reported by the authors, not a leaderboard entry
Not reported on 12 benchmarks
Wrong or missing number? Report an error