VLA GalleryAll modelsGallery

VLA Gallery / All models / Industry models

GigaBrain-0

GigaAI · 2025-10

VLAIndustry modelsNot disclosedflow matching

Reduces dependence on real-robot data with generated data from the GigaWorld world model (Real2Real, Sim2Real, human video conversion, etc.).

View architecture diagram in the galleryPaper ↗

Paper

GigaBrain-0: A World Model-Powered Vision-Language-Action ModelarXiv 2510.19430

Overview

Category
Industry models
Type
VLA
Released
2025-10
Organization
GigaAI
Open weights
Yes
Parameters
Not disclosed (GigaBrain-0-Small 402M)

Architecture

Backbone LLM
Gemma 2 · PaliGemma2 config (size not disclosed)
Vision encoder
SigLIP (RGB-D)
Action head
flow matching DiT (MoT) (flow matching)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
PaliGemma2 VLM + flow-matching DiT, mixture-of-transformers; RGB-D input (SigLIP depth channel zero-initialized); embodied CoT (10 2D trajectory keypoints, discrete action tokens, subgoal language)

Training

Pretraining accelerators
Not disclosed
Pretraining duration
Not disclosed
Pretraining steps & batch
Not disclosed
GPU/TPU-hours
Not disclosed
Fine-tuning
GPU and time not disclosed; laundry folding G1 300 demos · batch 128 · 40K steps; paper towel PiPER 100 demos · batch 128 · 20K steps
Data type
Real robot + simulation·synthetic (GigaWorld-generated) + public robot data
Data scale
In-house real robot 1,182 h (Agilex Cobot Magic 199 h, AgiBot G1 983 h; 14 environments, 3,100 m²); amount of public and generated data not disclosed
Main datasets
In-house data, AgiBotWorld, RoboMind, Open X-Embodiment, GigaWorld-generated data
Tasks & evaluation
Laundry folding, paper towel prep, juice making, table clearing, moving boxes·baskets, appearance·layout·viewpoint generalization; vs π0 laundry folding +30%, paper towel +10% (main text)
Robot
AgiBot G1, PiPER arm, Agilex Cobot Magic

Inference

Inference device & speed
On Jetson AGX Orin, GigaBrain-0-Small 0.13 s vs π0 1.28 s
Inference memory
Not disclosed
Weights bf16 (calc.)
Not disclosed
Where it runs
Onboard (Jetson AGX Orin, Small)

Reinforcement learning & base

Reinforcement learning
Not used · Supervised learning on real robot + generated data
Base control
Wheeled (Direct base velocity/pose output) · In mobile manipulation tasks (moving boxes·laundry baskets, AgiBot G1) the model does obstacle-avoiding movement·base repositioning; base command format not stated; PiPER bimanual arms are fixed

Notes

Key features
Reduces dependence on real-robot data with generated data from the GigaWorld world model (Real2Real, Sim2Real, human video conversion, etc.). The lightweight GigaBrain-0-Small (402M) runs in 0.13 s on Jetson AGX Orin.
Notes
Parameters (B) is null because the total for this model is not disclosed (only GigaBrain-0-Small 402M is disclosed). Per-task absolute success rates from figure bar labels could not be confirmed in verification, so only the differences in the text (+30%, +10%) are used. Weights released 2025-11-27 (repository Apache-2.0). GigaBrain-0.7 in the README is a 2026 release and excluded. Data hours (h) cover in-house data only. arXiv v1 2025-10-22.
Verification
Confirmed

Other models in this category