← Leaderboard

GO-1 (Genie Operator-1)

VLA
AGIBOT · Mar 10, 2025

AGIBOT’s first generalist policy, a Vision-Language-Latent-Action (ViLLA) model that adds a latent-action planner between an InternVL2.5-2B backbone and an action expert, pretrained on the AgiBot World dataset. Results are mostly self-reported on AGIBOT’s own real-robot tasks (the paper reports over 60% success on complex tasks and a 32% gain over RDT). Weights, plus a lighter GO-1 Air variant without the latent planner, were open-sourced on 2025-09-19.

Type
VLA
Parameters
3B
Backbone
InternVL2.5-2B + latent planner + action expert (ViLLA)
Weights
Open
License
CC BY-NC-SA 4.0
Commercial use
Not allowed
Released
Mar 10, 2025
Robots
AgiBot G1, AgileX Cobot Magic (ALOHA), Dual Franka
SO-101
–
LeRobot
–
Inputs
RGB · P · L
VRAM
~7GB

Benchmark results

1 of 13 benchmarks
#48 of 540.5%
Not reported on 12 benchmarks
Wrong or missing number? Report an error