DEV Community

武乐丹
武乐丹

Posted on

Gemini Robotics 2: DeepMind Just Gave Robots a Whole Body — VLA Breakdown for Developers

Gemini Robotics 2: DeepMind Just Gave Robots a Whole Body

On July 30, 2026, Google DeepMind unveiled Gemini Robotics 2 — and it may be the closest thing yet to a "general-purpose robot brain."

For years, vision-language-action (VLA) models like RT-2, Octo, and π0 focused on one thing: tabletop manipulation. A robotic arm. A gripper. Pick-and-place tasks on a fixed surface. Useful, but a long way from robots that can actually move through human spaces.

Gemini Robotics 2 changes that. It's the first VLA family that can control a full humanoid body — feet, legs, torso, arms, hands, fingers — in a single session.

Here's the honest breakdown: what's genuinely new, what's still locked behind partnerships, and who should care.

Three models, three capability tiers

Gemini Robotics 2 isn't one model — it's three, each targeting a different part of the robotics stack.

1. Gemini Robotics 2 (VLA) — the "doer"

The vision-language-action model that converts camera input and text instructions directly into motor commands. No intermediate motion-planning layer. It handles:

  • Full humanoid control — walking, crouching, stretching, reaching
  • 22-degree-of-freedom hand control (the SharkWave hand on Apptronik's Apollo 2)
  • Fine dexterity like tying knots and sealing ziplock bags
  • Standard gripper operation via the Franka Duo bi-arm platform

2. Gemini Robotics ER 2 (Embodied Reasoning) — the "thinker"

A vision-language model that acts as the robot's high-level brain. It plans multi-step tasks spanning minutes, communicates with humans, monitors progress, and re-plans when a step fails. It can also coordinate multiple robots working on the same task.

Crucially, ER 2 is available right now in Google AI Studio — you can test the reasoning model without owning a robot.

3. Gemini Robotics On-Device 2 — the "edge brain"

An efficient VLA that runs locally on robotic hardware with no cloud dependency. Its standout trait: embodiment transfer in hours. DeepMind says new robot bodies can be adapted with just a few hours of data and under 200 examples.

Whole-body control is the real headline

This is the part that matters. Previous VLA models controlled arms on tables. Gemini Robotics 2 controls an entire humanoid — and the difference shows in the demos:

  • "Put the watering can into the green bin on the bottom shelf" → Apollo 2 walks over, picks it up, walks to the shelf, crouches, and places it precisely.
  • "Clean up this cluttered room" → navigates obstacles, picks up objects at various heights, places them in designated containers.
  • Tying knots with a five-fingered hand → delicate manipulation that was previously impossible for AI-controlled robots.

Whole-body control forces the model to reason about balance, reach, trajectory, force, and object interaction simultaneously. That's a coordination problem neural networks have struggled with for years — and it's the difference between a robot that lives in a factory cage and one that can work in your house.

Multi-robot collaboration, unified reasoning

Another first: different robot types working together under one reasoning layer. DeepMind's demo shows a wheeled robot fetching objects while a humanoid handles fine placement on high shelves, coordinated through ER 2.

That matters because real environments are heterogeneous. A warehouse has floor bots, arm robots, mobile manipulators. A single reasoning layer that can task them cooperatively is the path to practical deployment — not a fleet of siloed, single-purpose machines.

Honest assessment: what's still missing

Capability Gemini Robotics 1.5 Gemini Robotics 2
Body control Upper body only Full body
Task duration <1 minute Several minutes
Self-correction Limited Full re-planning
Multi-robot Not supported Supported (ER 2)
Embodiment transfer ~1 week <1 day, <200 examples
Hand dexterity Basic grasp Knots, ziplock, precision
On-device inference No Yes

The limitations are real:

  • Access: VLA and On-Device models are restricted to early-access partners. Only ER 2 is publicly testable.
  • Speed: Humanoids still move slower than people — fine for structured environments, not yet for chaotic human spaces.
  • Cost: Apptronik Apollo 2 runs $50K+, Franka Duo $30K+, plus edge compute. This is enterprise hardware, not hobbyist territory.

Verdict: 8.8/10 — a genuine breakthrough, still early-stage

Gemini Robotics 2 delivers the three hardest things in robotics — generalization across embodiments, fine motor control, and long-horizon planning — in a unified model family. For research labs and enterprise robotics teams, that's transformative.

For the rest of us, the ER 2 reasoning model in AI Studio is a genuine taste of what's coming. As hardware costs fall and the models open up, this could be remembered as the moment "general purpose robots" went from sci-fi to shipping reality.


This review is part of our ongoing hands-on coverage of AI developer tools at AIPlaybook. We test what we write about — no affiliate bias, no sponsored verdicts.

Top comments (0)