Research Waitlist for early access; enterprise and research partnerships (pricing not published)

Gemini Robotics 2

Google DeepMind's robotics model layer for whole-body control, dexterous manipulation, and robots that coordinate with each other on physical tasks.

Updated 2026-08-02

8.2

AI Score / 10

Visit Gemini Robotics 2

Overview

Gemini Robotics 2 is Google DeepMind's robotics model layer, announced July 30, 2026. It takes a robot's camera and sensor input plus a plain-language instruction and produces the motor behaviour to carry it out — with the headline addition being whole-body intelligence: coordinating a base, torso, arms and hands as one system instead of treating manipulation as an isolated arm-planning problem. The launch demos leaned on messy household chores, tidying a garage and kitchen work, which is a deliberate choice: those tasks punish models that can only handle a clean tabletop and a single fixed viewpoint.

The other new claim is robot-to-robot collaboration — multiple machines working the same task rather than one robot executing a script in isolation. That is the harder problem, and it is where the demo footage did most of the persuading in the DeepMind X thread that carried the announcement past 100k views. This sits on top of the lineage DeepMind has been building since the original Gemini Robotics and the ER (embodied reasoning) variants: a vision-language model doing the thinking, a vision-language-action model doing the moving, and the ability to carry learned behaviour across different robot bodies.

This is not a product you sign up for and use. It is a model available through a waitlist and partner programme, aimed at robotics labs, hardware makers and research groups with a physical platform to put it on. If you don't have a robot, there is nothing here to try. Notably, DeepMind has not published pricing, a supported-platform list, or head-to-head benchmark figures alongside the launch post — so the honest read is that the demos are impressive and the reproducibility story is still missing. Treat the capability claims as unverified until third parties get hardware access.

Key features

Whole-body control

Coordinates base, torso, arms and hands as a single system rather than planning arm motion in isolation, which is what lets a robot reach into a cluttered space or shift its stance to complete a task instead of failing when the object is out of arm range.

Dexterous manipulation

Targets fine two-handed work — the grasping, reorienting and in-hand adjustment that household and workshop tasks demand. This has historically been the bottleneck between a robot that can point at an object and one that can actually pick it up and use it.

Multi-robot collaboration

Robots can work the same task together rather than each running an isolated policy. This is the genuinely new piece versus earlier Gemini Robotics releases and the hardest to fake in a demo video.

Language-instructed tasks

Takes plain-language instructions and turns them into physical behaviour, building on the embodied-reasoning approach in DeepMind's earlier robotics models where a VLM handles planning and a separate action model handles execution.

Pricing

Plan	Price	What's included
Early access waitlist	Not published	Sign up via the DeepMind announcement page. Access is gated and selective — no self-serve API key.
Enterprise / research partnership	Not published	Direct partnership route for robotics companies and research labs with hardware platforms. Terms negotiated, nothing listed publicly.

Early access waitlist Not published

Enterprise / research partnership Not published

Direct partnership route for robotics companies and research labs with hardware platforms. Terms negotiated, nothing listed publicly.

Pros & cons

Pros

✓Whole-body control targets the actual failure mode of most robot demos — tasks that need the machine to move its body, not just its arm
✓Robot-to-robot collaboration is a real step past the single-policy-per-robot pattern of earlier releases
✓Demos were shot on genuinely messy real-world scenes (garages, kitchens) rather than sanitised tabletops
✓Builds on DeepMind's existing embodied-reasoning stack, so it inherits Gemini's language understanding for instruction-following
✓Backed by Google DeepMind's hardware partnerships, which matters for a model that is useless without a robot to run on

Cons

×No public pricing, no self-serve access — waitlist and partnerships only, so most people reading about it cannot try it
×DeepMind published demo footage but no head-to-head benchmark numbers or supported-platform list alongside the launch, making the claims hard to verify
×Requires a physical robot platform, which puts it out of reach of anyone without hardware and an integration budget
×Curated demo videos are weak evidence for reliability — success rates on unscripted tasks are unknown until outside labs get access

How it compares

Tool	Best for	Pricing	Score
Gemini Robotics 2	—	Waitlist for early access; enterprise and research partnerships (pricing not published)	8.2/10
Perplexity AI vs Perplexity AI →	Knowledge workers and researchers who want cited, synthesized answers instead of a list of links to click through.	Freemium	9.4/10
NotebookLM	Students, researchers, and professionals who need answers grounded strictly in the specific documents they upload.	Free	9.1/10
Inkling-Small vs Inkling-Small →	—	Free — open weights download; Tinker Playground access	8.8/10