Google DeepMind's Gemini Robotics 2 Brings Whole-Body Control to Humanoids

Google DeepMind has unveiled Gemini Robotics 2, the latest version of its vision-language-action (VLA) model, and the headline capability is a significant one for anyone watching humanoid deployment: intelligent whole-body control, paired with advanced dexterity and multi-robot collaboration.

From tabletop tasks to full-body coordination

Where earlier work focused on upper-body manipulation, Gemini Robotics 2 extends control across an entire humanoid platform. The model can control the five-fingered, 22 degree-of-freedom SharpaWave hand on Apptronik's Apollo 2 humanoid robot — the kind of end effector complexity that has historically demanded hand-tuned control stacks. It can also operate standard two-fingered parallel grippers on a Franka Duo platform for complex dexterous tasks, suggesting the model is not locked to a single hardware configuration.

"Gemini Robotics 2 enables robots to reason through every movement, unlocking a broad range of tasks."

That reasoning claim matters. DeepMind's framing is that finesse, not just strength or speed, is the barrier to real-world utility.

"To be genuinely useful in our homes and workplaces, robots need finesse."

A two-layer architecture: reasoning above, action below

Alongside the VLA, DeepMind released Gemini Robotics ER 2, an embodied reasoning model that acts as the company's agent layer — enabling robots to communicate with people and understand the physical world. ER 2 can execute complex multi-step tasks, self-correct if a step fails, and generalize to novel situations. Notably, the model now understands when tasks begin and end and can pinpoint the moment key events occur, a capability that underpins longer, more reliable task sequences.

The update also introduces multi-robot collaboration, allowing different types of robots to work together — a workflow dimension a single robot cannot address alone.

On-device inference for disconnected environments

For deployments where connectivity is unreliable or latency is unacceptable, Gemini Robotics On-Device 2 is optimized to run locally on robotic devices.

"Gemini Robotics 2 can run locally on-device while adapting to entirely new robotic bodies in just a few hours."

The on-device model can adapt to new bi-arm robot embodiments with less than 200 examples — a data-efficiency figure that, if it holds in the field, meaningfully lowers the integration cost of porting the model across platforms.

Safety as a stated foundation

"Safety is foundational to its robotics research."

DeepMind introduced ASIMOV-Agentic, a benchmark for agentic safety orchestration and uncertainty resolution. On the practical side, Gemini Robotics ER 2 can detect when humans are nearby, trigger safety tool calls, and bring the robot to a safe stop if someone approaches too closely — behavior directly relevant to collaborative operation requirements.

Gemini Robotics ER 2 is available on Google AI Studio and in private preview on Gemini Enterprise Agent Platform, with the VLA and On-Device models offered to early-access partners.