VLA GalleryAll modelsGallery

VLA Gallery / All models / New in 2026

Gemini Robotics 2

Google DeepMind · 2026-07

VLANew in 2026Not disclosedNot disclosed

A VLA that converts visual·language input into whole-body motor control from toes to fingertips.

View architecture diagram in the galleryReport ↗Official page ↗

Paper

Gemini Robotics 2: Safety EvaluationsReport

Overview

Category
New in 2026
Type
VLA
Released
2026-07
Organization
Google DeepMind
Open weights
No
Parameters
Not disclosed

Architecture

Backbone LLM
Based on Gemini (size not disclosed)
Vision encoder
—
Action head
ER 2 planning → VLA execution (Not disclosed)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
Not disclosed; system 2/system 1 structure in which Gemini Robotics ER 2 (embodied reasoning VLM) splits into subtasks and the VLA executes

Training

Pretraining accelerators
Not disclosed
Pretraining duration
Not disclosed
Pretraining steps & batch
Not disclosed
GPU/TPU-hours
Not disclosed
Fine-tuning
Not disclosed
Data type
Not disclosed (same-family On-Device 2 card: image·text·robot sensor·action data)
Data scale
Not disclosed
Main datasets
Not disclosed
Tasks & evaluation
Apollo 2 + Inspire hands whole-body picking table 68.4% · floor 45.7% · shelf 76.3%; Apollo 2 + SharpaWave hands light bulb unscrewing 92% · screwing in 36%; Franka Duo pick-and-place 74.2% · tool kitting 78.9% · precision insertion 89.6%; human detection 99% · safe posture transition 96%
Robot
Apptronik Apollo 2 (Inspire hands, SharpaWave 22-DoF hands), Franka Duo/F3 Duo; across the whole family Dexmate · SO101 · Trossen, and Boston Dynamics Spot for ER 2

Inference

Inference device & speed
Not disclosed
Inference memory
Not disclosed
Weights bf16 (calc.)
Not disclosed
Where it runs
Not disclosed

Reinforcement learning & base

Reinforcement learning
Not disclosed · Training method details not disclosed
Base control
Humanoid legged (Direct whole-body joint/pose output) · VLA directly controls Apollo 2 whole body from toes to fingertips for walking·crouching (no separate locomotion controller mentioned); Franka Duo is fixed bimanual

Notes

Key features
A VLA that converts visual·language input into whole-body motor control from toes to fingertips. Covers humanoid whole-body manipulation and bimanual dexterous manipulation in one family.
Notes
Blog of 2026-07-30. Only a safety report, no technical report. ER 2 (embodied reasoning VLM, not a VLA) released via API the same day: progress classification 57.4%, moment finding 91.3%. Whether whole-body control uses a separate low-level controller is not disclosed. Introduces the ASIMOV-Agentic safety benchmark. Limited to early access partners.
Verification
Confirmed

Other models in this category