Embodied AI HubEmbodied AI Hub
ModelFeaturedPaper only

Gemini Robotics 2

Google DeepMind closed VLA for whole-body control — from bipedal balance to fingertip dexterity — plus sibling ER / On-Device models in the Gemini Robotics 2 family.

Official / sourcePaperBack to Model
VLA闭源Humanoid全身控制Google

Why we recommend it

Reference closed frontier for whole-body VLA capability; useful as a comparison baseline even when weights are not public.

Overview

Gemini Robotics 2 is Google DeepMind’s closed vision-language-action model for whole-body robot control. It sits alongside Gemini Robotics ER 2 (embodied reasoning / planning) and an On-Device 2 variant. Public demos emphasize feet-to-fingertips coordination; general access to VLA weights is limited.

Who it is for

Researchers and product teams tracking frontier closed VLAs, or comparing open baselines against DeepMind whole-body capability claims.

Key highlights

  • Closed frontier VLA for whole-body control
  • Family includes ER 2 (reasoning) and On-Device 2
  • Useful closed-source reference in open vs closed surveys

When to use

  • You are writing surveys / roadmaps that need a closed whole-body VLA reference
  • You are evaluating partner access to DeepMind robotics APIs

When not to use

  • You need downloadable open weights for local training
  • You require fully reproducible open benchmarks on your hardware

Getting started

  1. 1Read the DeepMind Gemini Robotics 2 model and blog pages.
  2. 2Separate VLA (motor) vs ER (planning) roles in the family.
  3. 3Compare claims against open baselines (OpenVLA, π0, GR00T N1.7, SmolVLA).

Papers & reading

How it compares

Caveats & pitfalls

  • VLA access is often waitlist / partner gated.
  • Parameter counts and training data mixes are not fully public.

Content reviewed 2026-08-02