Gemini Robotics 2 Expands Whole-Body Control

Google DeepMind’s update moves beyond tabletop arms. Public access remains limited.

What Changed

Google DeepMind unveiled Gemini Robotics 2 last week, according to The Robot Report.

The update adds whole-body control, higher hand skill, and multi-robot work. DeepMind says the system can turn a spoken goal into body motions.

Earlier models focused on a humanoid’s upper body and tabletop work. The new model is meant to guide walking, crouching, reaching, and object handling.

DeepMind also introduced Gemini Robotics ER 2. This model reads user requests, watches the setting, plans steps, and tracks task progress.

The Robot Report says ER 2 can help robots retry a failed step. It may adjust a hand motion rather than restart a full task.

Dexterity Claims

DeepMind says Gemini Robotics 2 can control Apptronik’s Apollo 2 humanoid. That includes the five-fingered SharpaWave hand with 22 degrees of freedom.

Google says the system can handle tasks such as knot tying and sealing a ziplock bag. It also supports two-finger grippers on a Franka Duo platform.

These are meaningful hand-control targets. Yet they remain company claims from demonstrations.

Access and Deployment Reality

Ars Technica reports that Gemini Robotics 2 includes three models. Only one is publicly available now.

ER 2 is available through Google AI Studio. The main action model and its on-device version are limited to early-access partners.

This is not broad commercial humanoid deployment. It is a developer release and partner-testing stage.

Key unknowns remain. The supplied evidence does not show uptime, task success rates, safety test results, operating cost, or deployment scale.