Google DeepMind Launches Gemini Robotics 2 With Whole-Body AI for Robots

Google DeepMind has introduced Gemini Robotics 2, a new generation of AI models designed to give robots broader physical capabilities, including full-body movement, improved dexterity, multi-step reasoning, and the ability to collaborate with other robots. The release expands the company's robotics platform beyond tabletop manipulation, with models that can control entire humanoid robots, operate locally on robotic hardware, and coordinate complex tasks over extended periods.

The announcement includes three new models, each targeting a different part of the robotics stack. Gemini Robotics 2 is Google's flagship vision-language-action (VLA) model that converts visual and language inputs into physical movements. Gemini Robotics ER 2 serves as the system's embodied reasoning model, enabling robots to understand instructions, communicate with humans, plan multi-step tasks, and monitor progress. Gemini Robotics On-Device 2 is an optimized VLA model designed to run directly on robotic hardware without relying on cloud connectivity.

Google said Gemini Robotics ER 2 is now available through Google AI Studio and is in private preview on the Gemini Enterprise Agent Platform. The Gemini Robotics 2 and On-Device 2 models are being made available to early-access partners.

A major focus of the update is whole-body robot control. While previous Gemini robotics models concentrated primarily on upper-body manipulation, Gemini Robotics 2 can now control an entire humanoid robot, allowing it to walk, crouch, stretch, and interact with objects while navigating real-world environments. Google demonstrated the model controlling Apptronik's Apollo 2 humanoid robot, completing tasks that required moving through a room before placing objects in designated locations.

The company also expanded the platform's manipulation capabilities across different robotic hardware. Gemini Robotics 2 can control the five-fingered SharpaWave hand on Apollo 2 to perform more delicate actions such as tying knots and sealing ziplock bags. It also supports standard parallel grippers on the Franka Duo platform for precision tasks including packing and insertion.

Google acknowledged that more progress is still needed before robots achieve human-level dexterity, noting that multifinger manipulation remains one of the more challenging areas.

The reasoning layer behind the system also received significant upgrades. Gemini Robotics ER 2 enables robots to interpret user requests, observe their surroundings, create execution plans, coordinate actions with the VLA model, and monitor whether tasks are completed successfully. Google said the updated model can execute task sequences lasting several minutes and involving hundreds of decisions while recognizing when individual steps succeed or fail.

According to the company, the model can also recover from mistakes during execution by retrying individual actions instead of restarting an entire workflow. This allows robots to adapt when objects move unexpectedly or when environmental conditions change during a task.

Another addition is multi-robot collaboration. Google said Gemini Robotics ER 2 allows different robot types to communicate with one another and divide work across a shared task, enabling workflows that would be difficult for a single robot to complete independently.

For applications that require low latency or operate without internet access, Google introduced Gemini Robotics On-Device 2. The model runs locally on robotic hardware and can be adapted to new dual-arm robot designs using only a few hours of training data. Google said adaptation typically requires fewer than 200 examples and can transfer to robots with different physical designs, sensors, and degrees of freedom.

The company also outlined new safety measures accompanying the release. Google introduced ASIMOV-Agentic, a benchmark intended to evaluate how embodied reasoning models handle uncertainty and safety decisions. The benchmark measures capabilities such as refusing unsafe tool requests, determining when a task cannot be completed safely, and requesting human intervention when appropriate.

Google said Gemini Robotics ER 2 also improves human proximity awareness, allowing robots to detect nearby people, trigger safety mechanisms, and stop movement when someone approaches too closely.

The company positions Gemini Robotics 2 as another step toward creating more general-purpose robotic intelligence capable of working alongside humans rather than performing only narrowly programmed tasks. While the demonstrations show advances in whole-body control, reasoning, dexterity, and collaboration, Google noted that movement speed, precision, and human-level manipulation remain areas of ongoing development.

See more here:

This analysis is based on reporting from Google.

Image courtesy of Google.

This article was generated with AI assistance and reviewed for accuracy and quality.