Google DeepMind Launch ‘Gemini Robotics 2’

Google DeepMind has unveiled Gemini Robotics 2, its next-generation artificial intelligence model designed to give robots more advanced reasoning, mobility, and dexterity. The latest system expands beyond controlling robotic arms to enabling full-body humanoid movement, allowing robots to walk, crouch, manipulate objects, collaborate with other robots, and carry out complex, multi-step tasks in real-world environments. The announcement marks one of Google’s biggest advances toward creating general-purpose AI-powered robots capable of operating safely alongside humans.

Gemini Robotics 2 introduces a family of models tailored for robotics, including a vision-language-action model for physical control, an embodied reasoning model for long-duration planning, and an on-device version that can run locally without cloud connectivity. Google demonstrated the technology on Apptronik’s Apollo 2 humanoid robot, showcasing tasks such as walking, picking up objects, tying trash bags, sealing bags, and organizing shelves.

Google DeepMind Introduces Gemini Robotics 2

Gemini Robotics 2 is designed to bridge the gap between AI reasoning and physical action.

The new platform enables robots to:

• Control their entire body from head to toe.
• Understand spoken and visual instructions.
• Perform multi-step tasks over extended periods.
• Recover from mistakes during execution.
• Interact naturally with people.
• Coordinate with multiple robots simultaneously.

Gemini Robotics 2 Family

Model	Primary Function
Gemini Robotics 2	Vision-language-action model controlling full-body robot movements
Gemini Robotics ER 2	Embodied reasoning and long-horizon task planning
Gemini Robotics On-Device 2	Local AI model that runs directly on robots without internet access

Full-Body Control Marks a Major Upgrade

Unlike earlier versions that primarily controlled robotic arms and upper-body movements, Gemini Robotics 2 enables coordinated movement across the robot’s entire body.

Demonstrated capabilities include:

• Walking and maintaining balance.
• Crouching and stretching.
• Picking up and placing objects.
• Operating five-fingered robotic hands.
• Manipulating everyday items with greater precision.

Google says the upgrade allows robots to perform more natural, coordinated movements needed for household, warehouse, and industrial environments.

Improved Dexterity and Embodied Reasoning

The system combines advanced perception, reasoning, and motor control to help robots understand their surroundings before taking action.

Examples demonstrated by Google include:

• Tying trash bags.
• Unscrewing and replacing light bulbs.
• Organizing shelves.
• Selecting specific objects from cluttered environments.
• Following spoken instructions while adapting to changing conditions.

Gemini Robotics ER 2 serves as the planning engine, enabling robots to reason through lengthy workflows, monitor progress, and adjust their actions if conditions change.

New Capabilities

Capability	Benefit
Full-body movement	Greater mobility and flexibility
Advanced dexterity	Better handling of delicate objects
Multi-step reasoning	Completion of longer, more complex tasks
Multi-robot collaboration	Coordinated execution across several robots
Natural language interaction	Easier communication with human users

On-Device AI Reduces Dependence on the Cloud

Google also introduced Gemini Robotics On-Device 2, which operates directly on robotic hardware.

Key advantages include:

• Offline operation without internet connectivity.
• Lower latency.
• Faster response times.
• Adaptation to different robot designs within hours.
• Greater reliability in industrial environments.

Running AI locally is expected to make robots more suitable for factories, warehouses, healthcare facilities, and remote locations where continuous cloud access may not be available.

Safety Remains a Core Focus

Alongside new robotics capabilities, Google introduced enhanced safety systems designed to reduce operational risks.

Safety features include:

• Human proximity detection.
• Automatic emergency stopping.
• Ability to reject unsafe commands.
• Requests for human assistance when uncertain.
• New benchmarks for evaluating AI-controlled robot safety.

Google says these safeguards are intended to support responsible deployment as robots become increasingly autonomous.

Intensifying Competition in Physical AI

The launch comes as technology companies race to develop general-purpose humanoid robots powered by foundation AI models.

Major competitors include:

• Tesla’s Optimus.
• Figure AI.
• Apptronik.
• Agility Robotics.
• Boston Dynamics.

Google’s latest release strengthens its position in robotics by combining Gemini’s multimodal intelligence with advanced physical control, an area viewed by many researchers as a key step toward “physical AI.”

Looking Ahead

Gemini Robotics 2 represents a significant milestone in Google DeepMind’s effort to bring advanced AI from digital applications into the physical world. By combining full-body robot control, embodied reasoning, on-device intelligence, and improved safety mechanisms, the new platform moves beyond isolated robotic demonstrations toward systems capable of completing practical, real-world tasks. While the technology remains in early-access testing, it highlights the rapid progress being made in autonomous robotics powered by multimodal AI.

Looking ahead, Google is expected to continue expanding Gemini Robotics through partnerships with robot manufacturers and enterprise developers. As AI models become more capable of understanding both language and physical environments, robotics could play an increasingly important role across manufacturing, logistics, healthcare, retail, and home assistance, potentially reshaping how intelligent machines interact with people and perform everyday work.