Google DeepMind's Gemini Robotics 2 expands AI control to robot full-body movement

Google DeepMind has unveiled the next generation of its robotics AI system, moving beyond hand and arm control to govern an entire robot's physical capabilities. The updated Gemini Robotics 2 extends machine learning influence across humanoid platforms, allowing coordinated motion from feet through fingertips—a significant leap from prior iterations that concentrated mainly on upper-body manipulation.

Whole-body coordination for complex tasks#

The expanded system now orchestrates intricate movements requiring full-frame coordination. Robots powered by Gemini Robotics 2 can navigate their environment by walking, bending, reaching, and crouching while simultaneously managing object interaction. Demonstrations reveal Apptronik's Apollo 2 model executing household tasks such as retrieving items from shelves and transporting containers—activities that previously demanded narrower, constrained robotic behavior.

The model also enhances fine-motor control by managing more sophisticated hand designs, including five-fingered configurations. This dexterity upgrade unlocks precision tasks: sealing bags, tying closures, or removing fixtures become feasible within a robot's action repertoire.

Enhanced reasoning and safety features#

Alongside physical control improvements, Google DeepMind refined Gemini Robotics ER, a vision-intelligence system that interprets settings and processes directives. The updated ER 2 version strengthens multi-step task execution and now discerns when operations commence and conclude—enabling robots to manage sustained assignments without repeated guidance.

The team also prioritized safeguarding. The latest iteration can detect nearby humans with greater precision, activate protective protocols, and halt motion if someone approaches dangerously, positioning it as DeepMind's "safest robotics model to date."

Multi-robot collaboration and local processing#

A particularly notable capability allows diverse robot types to communicate and cooperate on shared objectives. Footage demonstrates one Apollo 2 unit directing a separate dual-arm platform to organize tools—a preview of heterogeneous robot teams tackling coordinated work.

On-device processing improvements let robots adopt new physical architectures—different shapes, sensor arrays, and movement capabilities—more rapidly, enabling faster deployment across varied platforms without internet reliance.