Google unveils Gemini Robotics 2 with whole‑body AI control

Google announced Gemini Robotics 2, its newest AI layer that lets robots control whole bodies—from feet to fingertips—while running locally on‑device and collaborating with other robots [1].

Whole‑body control and dexterity

Gemini Robotics 2’s vision‑language‑action (VLA) model can command full humanoid robots, enabling actions such as walking, crouching, stretching and manipulating objects in cluttered environments [1]. In a demo with Apptronik’s Apollo 2 humanoid, the system was asked to “put the watering can into the green bin on the bottom shelf,” and the robot walked, grasped and placed the can precisely [1]. The same model also drives the 22‑degree‑of‑freedom SharpaWave hand to tie knots or seal zip‑lock bags, and operates standard two‑finger grippers on a Franka Duo platform for tight packing tasks [1]. These capabilities extend beyond prior Gemini versions, which were limited to upper‑body motions and table‑top tasks.

Reasoning, collaboration and on‑device efficiency

The embodied reasoning (ER) model, Gemini Robotics ER 2, serves as a high‑level brain that parses user instructions, plans multi‑step actions lasting several minutes, and coordinates with the VLA model to execute them [1]. It now tracks task boundaries and can self‑correct when steps fail, a step up from earlier models that handled only short, repetitive sequences. Multi‑robot collaboration is also introduced, allowing different robot types to communicate and divide labor on complex workflows [1].

Gemini Robotics On‑Device 2 is a lightweight VLA optimized for local execution, enabling adaptation to new robot embodiments in a few hours with fewer than 200 training examples [1]. This addresses latency and connectivity constraints that have limited many field deployments.

Safety and the path to general‑purpose physical AI

Google emphasizes safety, adding the ASIMOV‑Agentic benchmark to evaluate the model’s ability to refuse unsafe tool calls and request human intervention when uncertain [1]. The ER 2 model reportedly improves proximity detection and can trigger safe stops if a human approaches too closely, aligning with collaborative safety standards [1]. These safety layers are positioned as essential steps toward the broader goal of general‑purpose physical AI, moving beyond single‑task automation [1].

What to watch

• Rollout timeline – Early‑access partners will receive the VLA and On‑Device models this quarter; broader availability remains unspecified.
• Enterprise adoption – Gemini Robotics ER 2 is in private preview on the Gemini Enterprise Agent Platform; uptake by large manufacturers could signal market traction.
• Competitive response – Rivals such as Boston Dynamics and Nvidia’s Isaac platform may accelerate their own whole‑body and on‑device AI offerings to match Google’s capabilities.

Google’s Gemini Robotics 2 demonstrates a tangible shift from narrowly programmed robots to systems that can reason, adapt and collaborate across whole bodies, narrowing the gap between current automation and the vision of general‑purpose physical AI. The open question is whether the on‑device adaptation speed and safety guarantees will be enough to drive widespread commercial deployment.