Gemini Robotics 2 brings whole body intelligence to robots

From feet to fingertips — we are school robots intelligent whole-body control, good dexterity, and teamwork to complete a wide scope of analyzable tasks

For decades, we’ve dreamed of robots that tin seamlessly measurement into our world and lend a hand. Now, that imagination takes a important stride forward.

Most robots are pre-programmed aliases teleoperated for narrow, repetitive task sequences. They deficiency the expertise to genuinely study for themselves aliases accommodate to unpredictable environments. Moreover, transferring learned skills from 1 robot assemblage to different remains incredibly difficult. To return connected the hardest problems astatine scale, robots of each style and size request AI models giving them the expertise to think, act, and interact intelligently to safely complete tasks.

We demonstrated really Gemini's multimodal knowing could thrust real-world action pinch Gemini Robotics. Today, we are introducing Gemini Robotics 2 - the intelligence furniture powering the adjacent procreation of genuinely adaptable robots. As it takes its first literal steps, this awesome beforehand unlocks intelligent whole-body control, precocious dexterity, and multi-robot collaboration.

Gemini Robotics 2 enables robots to logic done each movement, unlocking a wide scope of tasks. For example, it tin alteration a humanoid to walk, crouch, stretch, and manipulate objects to cleanable up a cluttered room. It tin moreover squad up pinch different robots to decorativeness the occupation faster. And this profound intelligence tin besides tally locally on-device while seamlessly adapting to wholly caller robotic bodies successful conscionable a fewer hours.

We are making this imaginable done 3 highly tin models:

• Gemini Robotics 2: Our astir precocious vision-language-action exemplary (VLA) that converts imagination and connection input into centrifugal control, enabling a robot to return action. This exemplary is tin of controlling afloat humanoids, from feet to fingertips, and different bi-arm robots. It besides brings a caller level of dexterous manipulation connected some hands and grippers.
• Gemini Robotics ER 2: Our astir tin embodied reasoning (ER) model. It is simply a imagination connection exemplary (VLM) that acts arsenic our agent, enabling robots to pass pinch humans, understand the beingness world and scheme multi-step tasks lasting respective minutes. We are besides introducing the expertise for robots to activity together arsenic a team.
• Gemini Robotics On-Device 2: Our astir businesslike vision-language-action exemplary (VLA) optimized to tally locally connected robotic devices. This exemplary tin now execute accelerated adjustment to wholly caller robot embodiments pinch a fewer hours of data.

Gemini Robotics ER 2, our reasoning model, is now disposable connected Google AI Studio and successful backstage preview connected Gemini Enterprise Agent Platform. Our VLA and On-Device models are disposable to early-access partners. Read really to bring these models to your hardware connected our Developer blog.

Humanoids successful motion: Managing whole-body tasks

The world is built for quality movements; it requires america to reach, bend, and equilibrium successful tight, cluttered spaces. While our erstwhile models controlled the humanoid’s upper-body to execute table-top tasks, Gemini Robotics 2 expands beingness AI into whole-body motions.

For the first time, our exemplary tin now power full humanoid robots, translating intent into intelligent whole-body control. For example, erstwhile controlling Apptronik’s Apollo 2 humanoid robot, we tin inquire it to “put the watering tin into the greenish bin successful the bottommost shelf.” Apollo processes the instruction, walks to the table, and picks up the watering can, takes a fewer steps to the shelves, and places it precisely successful its destination. While our robots person much to beforehand successful activity speed, this is an important measurement towards the skills needed to complete much complex, real-world tasks that require whole-body coordination.

Bringing precocious dexterity to hands and grippers

To beryllium genuinely useful successful our homes and workplaces, robots request finesse. Gemini Robotics 2 unlocks a caller level of beingness dexterity crossed different extremity effectors, whether a robot is utilizing hands aliases grippers, enabling robots to beryllium much useful than ever before.

The exemplary tin now power the five-fingered, 22 degree-of-freedom SharpaWave manus connected the Apollo 2 robot to complete delicate actions for illustration tying knots aliases sealing a ziplock bag. It tin besides run modular two-fingered parallel grippers connected a Franka Duo platform to execute analyzable dexterous tasks (e.g. tight packing). We are continuing to beforehand the level of precision and velocity to execute human-level dexterity.

Unlocking precocious tasks pinch agentic reasoning and multi-robot collaboration

Most real-world tasks require aggregate steps complete an extended play of time. To negociate this complexity, our embodied reasoning (ER) model, Gemini Robotics ER 2, serves arsenic the robot’s high-level brain, processing personification instructions and communicating pinch humans. It observes the room, reasons astir the steps needed to complete the task, coordinates pinch the VLA to transportation retired the actions, and tracks advancement until the task is done. This setup allows robots to execute analyzable multi-step tasks, self-correct if a measurement fails, and generalize to caller situations and goals.

In this update, we are enabling robots to much reliably execute longer task sequences, lasting respective minutes and involving hundreds of decisions. Gemini Robotics ER 2 now understands erstwhile tasks statesman and end, and tin pinpoint the infinitesimal cardinal events occur, marking a measurement alteration successful advancement understanding.

Furthermore, we are introducing multi-robot collaboration. This enables different types of robots to pass and activity together to lick analyzable workflows a azygous robot could not do alone.

Adapting accelerated on-device models for immoderate robot

Many robotic applications request to run without web latency aliases net connectivity. Gemini Robotics On-Device 2 is built specifically to grip these constraints — it is our most-efficient vision-language-action exemplary (VLA) optimized to tally locally connected robotic devices.

This exemplary is natively multi-embodiment and inherits our precocious “motion transfer” techniques from Gemini Robotics 1.5. We tin now accommodate to caller bi-arm robot embodiments pinch conscionable a fewer hours of adjustment time, typically pinch little than 200 examples. This useful moreover pinch caller embodiments pinch drastically different shapes, sensors and degrees of freedom, arsenic shown beneath pinch a divers group of tasks being performed by the Dexmate, SO101, and Trossen platforms.

Advancing our committedness to safe and responsible robotics

Safety is foundational to our robotics research. As robots summation much beingness capabilities, we are committed to ensuring end-to-end information and alignment. With each release, we’ve taken a multi-layered attack that combines accepted beingness information measures pinch robust AI information frameworks.

Gemini Robotics 2 specifically advances robotics information for navigating the uncertainty of the existent world and collaborating alongside humans.

We’re introducing ASIMOV-Agentic, a caller benchmark for agentic information orchestration and uncertainty resolution. For example, it measures the embodied reasoning agent’s expertise to garbage unsafe instrumentality calls from a VLA.It besides measures the agent’s expertise to foretell whether a task is imaginable and to proactively petition quality involution erstwhile uncertain.

Additionally, pinch enhanced embodied reasoning, Gemini Robotics ER 2 is our safest robotics exemplary to day successful information constraint pursuing and quality proximity benchmarks. It tin amended observe erstwhile humans are nearby, trigger information instrumentality calls and bring the robot to a safe extremity if personification approaches excessively closely. This is simply a cardinal request successful collaborative information standards. Read our Gemini Robotics 2: Safety Technical Report for much details.

Building towards general-purpose beingness AI

Gemini Robotics 2 marks an important milestone connected the way toward solving AGI successful the beingness world. Unlocking the existent imaginable of robotics requires moving past single-task automation toward general-purpose intelligence. By building this halfway intelligence, our extremity is to alteration AI successful the beingness world that tin activity alongside humans to lick analyzable challenges.