The next generation of robotic intelligence brings whole-body control, human-level dexterity, and multi-robot teamwork to the physical world.
- Whole-Body Mastery and Dexterity: Robots can now intelligently coordinate their entire bodies—from walking and crouching to delicately tying knots—moving beyond simple, repetitive tasks.
- Advanced Reasoning and Teamwork: A new embodied reasoning brain allows robots to plan complex, multi-step tasks lasting several minutes, self-correct errors, and even collaborate with other robots.
- Rapid, Safe On-Device Adaptation: Highly efficient models can adapt to entirely new robotic bodies in just a few hours without internet connectivity, all while adhering to strict, multi-layered safety protocols.
For decades, we’ve dreamed of a world where robots seamlessly step into our daily lives to lend a helping hand. Until now, that vision has been constrained by technological limits. Most robots are strictly pre-programmed or teleoperated to perform narrow, repetitive sequences. They lack the fundamental ability to learn, adapt to unpredictable environments, or transfer a learned skill from one mechanical body to another. To solve real-world problems at scale, robots need AI models that empower them to think, act, and interact intelligently.
Today, that vision takes a massive leap forward. Following the initial success of bringing multimodal understanding to physical hardware, we are entering a new era with Gemini Robotics 2. Serving as the foundational intelligence layer for the next generation of adaptable robots, this major advancement unlocks intelligent whole-body control, fine dexterity, and multi-robot collaboration.
The Intelligence Trio Powering the Future
To achieve this profound leap in physical AI, Gemini Robotics 2 relies on three highly capable, specialized models working in harmony.
| Model | Type | Core Capabilities & Availability |
|---|---|---|
| Gemini Robotics 2 | Vision-Language-Action (VLA) | Converts vision and language into motor control. Manages full humanoids from feet to fingertips, and advanced dexterity on bi-arm robots. Available to early-access partners. |
| Gemini Robotics ER 2 | Vision-Language Model (VLM) / Embodied Reasoning | The “high-level brain.” Plans multi-step tasks, communicates with humans, and coordinates multi-robot teams. Available on Google AI Studio & Gemini Enterprise Agent Platform. |
| Gemini Robotics On-Device 2 | Vision-Language-Action (VLA) | Highly efficient, optimized for local on-device processing. Adapts to entirely new robot embodiments in a few hours. Available to early-access partners. |
Humanoids in Motion: Mastering the Cluttered World
The physical world was built for human movement. It requires us to reach, bend, balance, and navigate tight, cluttered spaces. While previous models were restricted to controlling a humanoid’s upper body for tabletop tasks, Gemini Robotics 2 expands physical AI into comprehensive, whole-body motion.
For the first time, this model translates human intent into intelligent, full-body control. When paired with Apptronik’s Apollo 2 humanoid robot, you can simply ask it to “put the watering can into the green bin on the bottom shelf.” Apollo processes the command, walks to the table, picks up the watering can, takes a few precise steps to the shelves, and places the object exactly where it belongs. While movement speed will continue to advance, this represents a crucial milestone in equipping robots for complex, real-world coordination.
Furthermore, to be genuinely useful in our homes and workplaces, robots need finesse. Gemini Robotics 2 unlocks a new level of physical dexterity across various end effectors. Whether operating the five-fingered, 22-degree-of-freedom SharpaWave hand on the Apollo 2 to tie knots and seal ziplock bags, or using standard two-fingered parallel grippers on a Franka Duo platform for tight packing, robots can now manipulate objects with unprecedented precision.
Unlocking Teamwork and Agentic Reasoning
Real-world tasks rarely consist of a single action; they require multiple steps over an extended period. To manage this complexity, Gemini Robotics ER 2 serves as the robot’s executive brain. It observes the room, reasons through the necessary steps, coordinates with the VLA to execute actions, and continuously tracks progress.
This model marks a step-change in spatial and temporal understanding. It can reliably execute task sequences lasting several minutes, involving hundreds of autonomous decisions. It knows when tasks begin and end, can self-correct if a step fails, and generalizes its learning to novel situations. Most excitingly, ER 2 introduces multi-robot collaboration. Different types of robots can now communicate and work together to solve complex workflows that a single robot could never accomplish alone.
Fast Adaptation and Uncompromising Safety
Many robotic applications must operate in environments lacking internet connectivity or where network latency is unacceptable. Gemini Robotics On-Device 2 is built specifically for these constraints. Inheriting advanced “motion transfer” techniques from its predecessor, this natively multi-embodiment model can adapt to completely new bi-arm robots with drastically different shapes, sensors, and degrees of freedom in just a few hours—typically requiring fewer than 200 examples. This rapid adaptation has already been proven across diverse platforms like Dexmate, SO101, and Trossen.
As physical capabilities expand, safety remains foundational. Gemini Robotics 2 takes a multi-layered approach to ensure end-to-end alignment in unpredictable environments. To measure this, we are introducing ASIMOV-Agentic, a new benchmark designed to evaluate safety orchestration and uncertainty resolution. It rigorously tests an agent’s ability to refuse unsafe tool calls, predict task feasibility, and proactively request human intervention when uncertain.
Additionally, Gemini Robotics ER 2 is our safest model to date when it comes to human proximity. It can intelligently detect when humans are nearby, trigger safety tool calls, and bring the robot to a safe, complete stop if someone approaches too closely, aligning with strict collaborative safety standards.
Building Towards General-Purpose AI
Gemini Robotics 2 is not just an incremental update; it is a vital milestone on the path toward Artificial General Intelligence (AGI) in the physical world. By moving past rigid, single-task automation and building a core layer of general-purpose intelligence, we are moving closer to a future where AI safely and intuitively works alongside humans to solve our most complex physical challenges.
