Google DeepMind unveils Gemini Robotics 2 with intelligent whole-body control for humanoids

Google reveals Gemini Robotics 2.0, promising improved dexterity and safety

Robots powered by Google's Gemini AI models are now more capable. With the debut of Gemini Robotics 2, these physical bots can now accomplish more complex tasks, continuously analyze changing environments, and collaborate with other robots. This is thanks to a trio of new sub-models, one of which is publicly available for developers starting today. Videos of robots running, dancing, and backflipping have been a staple of the Internet for years, but these machines were programmed to perform these very narrow tasks. The goal of Gemini robotics is to create a generalist robot, one that can do anything a human could do. Google DeepMind scientists sometimes call this "physical AGI." Essentially, you tell a robot what to do, and it does it. With the 2.0 release, Google says its robotics AI can control an entire humanoid robot with improved dexterity, even for machines with complex humanoid hands. This starts with Gemini Robotics ER 2, an upgraded "embodied reasoning" model that DeepMind claims is a significant leap over the previous 1.6 release. It's integrated with the Gemini Live API, giving developers the opportunity to experience that supposed leap forward. Gemini Robotics ER 2 is what's known as a vision language model (VLM). It's designed to understand instructions and the world around it. The big upgrade here is that ER 2 can process live video feeds from the robot's cameras, allowing the system to track progress as the robot lumbers from one step to the next. Google notes that Gemini Robotics ER 2 can classify video frame completeness with almost 60 percent accuracy. That's still far from perfect, but it's much better than the 1.6 release or what you can get with the visual understanding of competing AI models. Finding specific moments in video feeds is also key to completing a task correctly. When you ask a robot to pour a cup of coffee, you definitely want it to know when to stop pouring. ER 2 apparently does this much better, identifying key moments with almost 90 percent accuracy. As the robots execute multi-step tasks, the embodied reasoning model allows them to understand failures in real time. The system can then attempt that single step again rather than going back to the start. For example, the robot can just readjust its hand position and motion if a ball it's trying to pick up rolls away or someone moves a container. The new embodied reasoning release is also what gives Google DeepMind's upgraded robot AI the ability to collaborate. The video demos show Apptronik's Apollo 2 and the simpler Franka F3 Duo working together on a task without getting in each other's way. While the test robots still can't match the speed or grace of a human, the video demos include plenty of real-time footage of the robots in action, and they do seem much less hesitant than they were in past tests. Understanding is only the first step -- getting the robot to move around in the physical world is the purview of another model. After mapping out the task, the vision language model hands things over to an upgraded vision-language-action known simply as Gemini Robotics 2. This AI model generates robot actions from those instructions in the same way other generative systems create text or images. There's also a low-latency offline version of this called Gemini Robotics On-Device 2. These models are currently limited to a small group of testers. Google says the new action models are much more accurate and efficient. Even the smaller on-device version can adapt to new robot designs with just a few hours of movement data, or around 200 examples. Heading off the robot apocalypse The issues with AI hallucinations are well known at this point, but the potential harm from mistakes when the AI has a physical embodiment sharing space with humans could be much greater. With each release of Gemini Robotics, Google DeepMind has stressed that it takes this risk seriously. According to DeepMind, each layer in Gemini Robotics includes "traditional physical safety measures with robust AI safety frameworks." With the release of Gemini Robotics 2, there's a new safety benchmark called ASIMOV-Agentic. This test evaluates models across a variety of safety factors. It can assess whether an embodied reasoning agent will refuse unsafe tool calls from a VLA. It can also determine whether a given task is possible to complete safely, as well as whether the model is able to call for human assistance if it's unsure about safety. Google DeepMind notes that Gemini Robotics ER 2 is the team's safest model yet, showing robust ability to understand safety and halt actions when a human is too close to the robot. Google's new safety benchmark is available in its entirety on Hugging Face.

Google's Gemini Can Now Stomp Around as a Humanoid Robot

Google DeepMind just released a new version of its artificial intelligence model Gemini, and it can control a range of different robots -- including humanoids capable of dextrous tasks like screwing in lightbulbs and tying trash bags. Gemini Robotics 2 combines several different AI models into a single system. Taken together, they allow a robot to make sense of its surroundings and how to act in it. A vision language model (VLM), which understands images and video, can communicate with humans and reason how to perform different tasks. Two vision language action (VLA) models, trained to understand how to move in physical space, control the robot's full-body movement as well as the movements of grippers or hands. In video demonstrations shared ahead of the release, the company showed several different robots performing complex tasks autonomously using the amalgamated model. In one demo, Apptronik's Apollo 2 robot used hands from a company called Sharpa to tidy shelves. Google DeepMind trained the model to perform these tasks using a mix of human teleoperation, video examples, and simulations -- it's not yet possible for AI models to perform a wide range of complex tasks without specific training. Although Anthropic and OpenAI have taken a lead with chatbots and AI coding tools, Google has a stronger track record in robotics research, and has published important work on using AI to train robots to do useful things. The release is another sign that the search giant is betting AI will need to break free from the digital realm to realize its full potential. (It previously partnered with Boston Dynamics, a leader in legged robots, to provide the brains for those machines.) "It's another milestone in our path towards really getting towards what we call like physical AGI, which means we get a robot to do anything that a human can," Carolina Parada, head of robotics at Google DeepMind, tells WIRED. Giving frontier AI models access to robots so that they can wander around workplaces or homes and manipulate objects does, however, come with risks. Previous research has shown that using frontier AI to control robots can produce unexpected and sometimes dangerous behavior. And the idea that these models can take sudden or unwanted actions in the digital realm became apparent recently, when an unreleased AI agent developed by OpenAI hacked several systems. "The safety question is even more pressing because you're putting them in a lot of other situations," Parada says. "There's a lot of uncertainty that will show up, and so you want to be able to understand the safety question more deeply." Parada says Google takes a multi-layered approach to safety, with guardrails applied on each model layer. It's also introducing ASIMOV-Agentic, a new benchmark for measuring the safety of various AI systems collaborating to control a robot. The benchmark detects whether a command will result in harmful or uncertain outcome. The company's CEO, Demis Hassabis, previously told WIRED that he hopes to develop an AI operating system for many different robots similar to the Android operating system for smartphones.

Google DeepMind's new AI model can control a robot's entire body

Google DeepMind says the latest version of its Gemini Robotics AI model can "control entire humanoid robots." While the previous model focused on controlling a humanoid robot's upper body, Gemini Robotics 2 now supports "whole-body motions" ranging from its feet to fingertips, according to an announcement on Thursday. The new model will allow humanoid robots to perform a wider range of actions, as it allows them to walk, crouch, stretch, and manipulate objects. Videos shared by Google show how Apptronik's Apollo 2 robot can bend over to pick up a watering can, as well as find and take specific items off a shelf. Though Google DeepMind notes that its robots "have more to advance in movement speed," it adds that this update "is an important step towards the skills needed to complete more complex, real-world tasks that require whole-body coordination." Additionally, Gemini Robotics 2 supports better dexterity, as it can now control more complex, five-fingered hands. That enables robots to perform tasks like sealing a Ziploc, tying a trash bag, or unscrewing a lightbulb. Google DeepMind is updating Gemini Robotics ER (embodied reasoning) as well, a vision-language model that helps robots to analyze their surroundings, process instructions, and perform multi-step tasks. Gemini Robotics ER 2 is better at completing tasks over an extended period of time and "now understands when tasks begin and end." Google DeepMind says this update also allows multiple robots of different types to work together and complete tasks, with one video showing how Apollo 2 instructs Google's dual-arm robot to put tools inside a bin while cleaning the garage. The company notes Gemini Robotics ER 2 is its "safest robotics model to date," as it can "better detect when humans are nearby, trigger safety tool calls and bring the robot to a safe stop if someone approaches too closely." Meanwhile, Google DeepMind has brought improvements to its Gemini Robotics On-Device Model, which can run locally on a robot without an internet connection. This model can now adapt to new embodiments faster, including those with "drastically different shapes, sensors and degrees of freedom."

Google's new Gemini Robotics 2 platform allows for 'intelligent whole-body control' - Engadget

The company made a video of the bots in action doing stuff like cleaning trash and picking up watering cans. Google DeepMind just announced a new version of its artificial intelligence model Gemini that can control a range of robots. This includes humanoid robots capable of tasks like cleaning up trash and picking up watering cans. The company released a video of the Gemini Robotics 2 platform in action and the results are fairly impressive. It shows robots putting a tape into a boombox, screwing in a lightbulb and even tying up a garbage bag. The original Gemini Robotics system could already power bots that performed finicky actions like closing storage bags and folding origami, but that was done via robotic arms and hands. This update brings "intelligent whole-body control" into the mix. I'm fairly skeptical of this stuff, but Google promises the aforementioned video features "real-time footage" of "fully autonomous" robots. This operates in stark contrast to Elon Musk's Optimus robots, as a high-profile demo was proven to have used teleoperators to control the bots. Google is doing this by stuffing the robots with several different AI models, which results in a multimodal understanding of their surroundings. There's a vision language model to help with understanding and two vision language action models that control full-body and hand movement. It's worth noting that these aren't robots aren't general-purpose devices that can do just about anything. The model was specifically trained to perform every task in the video, using a mix of human teleoperation, video examples and simulations. Google DeepMind says "it's another milestone in our path towards really getting towards what we call like physical AGI, which means we get a robot to do anything that a human can," according to a statement published by Wired. There is one final caveat here. These are robots filled with AI and, well, AI gets stuff wrong. That's frustrating when looking for an answer to a query, but potentially deadly when dealing with a heavy human-sized robot. Google says it's taking a multi-layered approach to ensure future robots are safe. Each model layer has guardrails and it has introduced a new benchmark called ASIMOV-Agentic for detecting whether a command will lead to a harmful outcome. "The safety question is even more pressing because you're putting them in a lot of other situations," Carolina Parada, head of robotics at Google DeepMind, told Wired. "There's a lot of uncertainty that will show up, and so you want to be able to understand the safety question more deeply." Finally, this is all early stages. Google isn't about to roll out consumer-facing robots anytime soon. Robotics rival Elon Musk, on the other hand, promises Tesla's upcoming Optimus robots will be "the biggest product ever" that will result in $10 trillion in sales. He once stated we'd have 50,000 to 100,000 Optimus robots by 2026 and 1,000 of the little bots roaming around manufacturing facilities by the end of 2025. The latter didn't happen and the former won't happen.

Google DeepMind's Gemini Robotics 2 controls whole humanoids

Google DeepMind has released Gemini Robotics 2, a family of models that can control a humanoid from its feet to its fingertips, coordinate several robots at once, and adapt to a new machine in a few hours. It is a real step toward "physical AI," though DeepMind's own figures show the robots are still slow and clumsy at the fiddliest tasks. Google DeepMind wants one AI brain to run every robot, and it just taught that brain to use its whole body. It has released Gemini Robotics 2, a family of models that can control a humanoid from its feet to its fingertips. The system can also coordinate several machines at once and adapt to a new robot body in a few hours. It is the clearest sign yet of what the industry calls "physical AI," and Wired framed it as a real step toward "physical AGI." The pitch is simple. The same kind of model that writes your emails is now learning to walk across a cluttered room and tidy it up. From arms to the whole body The headline change is control. DeepMind's earlier robot models mostly moved an upper body to do tabletop tasks. Gemini Robotics 2 drives the entire machine. It can make a humanoid walk, crouch, stretch and balance while manipulating objects in tight spaces built for humans. In a demo, Apptronik's Apollo 2 robot heard a single instruction: put the watering can in the green bin on the bottom shelf. It walked to a table, picked up the can, stepped over to the shelves, bent down and placed it. That sounds trivial. Coordinating legs, torso, arms and hands from one prompt is not. Dexterity has improved too. The model can drive Apollo's five-fingered, 22-joint hand to tie knots and seal a ziplock bag, and it can run simpler two-fingered grippers on other platforms. "Our goal is to bring AI into the physical world and then build the intelligence layer that can be used by every robot," said Carolina Parada, DeepMind's head of robotics. Three models, one system The release is really three models. Gemini Robotics 2 is the vision-language-action model that turns what a robot sees and hears into motor commands. It handles the physical doing. Gemini Robotics ER 2 is the reasoning layer, a high-level brain that plans multi-step jobs. In its developer briefing, Google showed it tracking its own progress on a live video feed. It can call tools like Google Search, and even steer a Boston Dynamics Spot robot to fetch a snack. It also lets different robots work as a team, a wheeled machine and a humanoid splitting a task. It is available now to developers through the Gemini API and Google AI Studio. The third, On-Device 2, runs locally with no internet connection. It can be fitted to an entirely new robot body with fewer than 200 examples and a few hours of training. That last point matters, because moving a learned skill from one machine to another has long been one of robotics' hardest problems. Impressive, and still slow DeepMind was unusually frank about the limits. Its own numbers show the gap. Bloomberg reported the system could unscrew a light bulb 92 per cent of the time, but fiddlier jobs lagged badly. The Chosun Daily put the trash-bag tie at 44 per cent and the ziplock seal at 40 per cent. The robots are also slow. They pause to think through moves a person makes without a second thought. Kanishka Rao, a DeepMind robotics director, said true dexterity remains a distant goal, and that robots still learn far less efficiently than humans, who adjust after one or two mistakes. This is a pattern across the field. Rival efforts from robotics foundation-model startups, and dexterity work on rival humanoids, keep hitting the same wall. The demos dazzle, but the machines are still far from ready for the home. Safety moves up the stack As robots gain the ability to move around people, DeepMind put more weight on safety. Gemini Robotics ER 2 is its safest model yet, it says, better at spotting when a person is close and halting until the area is clear. It resumes only when the space is empty again. The company also released a benchmark, ASIMOV-Agentic, that tests whether the reasoning model will refuse an unsafe command from the action model, and whether it flags when a task is impossible or asks a human for help. Naming a robot-safety benchmark after Isaac Asimov is on the nose. The underlying worry, machines acting on flawed instructions, is not. A hardware problem Google cannot solve There is an awkward backdrop. Google makes the software, not the robots, and the hardware supply is getting political. As Axios noted, the US just moved to ban future sales of Chinese-made robots on security grounds. Many of the bodies this software might run on are built in China. Google is working with Western partners including Apptronik, Boston Dynamics and Agile Robots, and more than 100 trusted testers. Its rivals are circling the same prize. OpenAI and Nvidia are both building robot models, and everyone is chasing the same idea: one model, any body. DeepMind is careful to call this a milestone, not the finish line. Gemini Robotics 2 makes robots more general and more useful than before. It also makes plain how far a machine still is from the easy competence of a human tidying a room.

Google's Gemini Robotics 2 gives humanoid robots full-body control

Google has introduced Gemini Robotics 2, a new robotics model designed to control entire humanoid robots, giving them the ability to walk, bend, balance, and manipulate objects while completing complex tasks. The company says the system can also coordinate multiple robots and adapt to different robot bodies with only a few hours of training. The release builds on Gemini Robotics introduced earlier this year, expanding capabilities beyond tabletop manipulation to full-body movement. Google is also launching two companion models: Gemini Robotics ER 2 for high-level reasoning and Gemini Robotics On-Device 2, which runs locally on robots without requiring cloud connectivity. Unlike conventional robots that are programmed for repetitive tasks, the new models are designed to reason through actions, adapt to unfamiliar environments, and carry out multi-step jobs with minimal retraining. Google said the vision-language-action model can control humanoid robots as well as dual-arm robotic systems, enabling them to carry out household and industrial tasks using both hands or robotic grippers. One of the biggest changes is whole-body control. While earlier models focused mainly on upper-body movements, Gemini Robotics 2 enables humanoids to walk across rooms, crouch, stretch and interact with objects. During demonstrations, Google's model controlled Apptronik's Apollo 2 humanoid robot, directing it to pick up a watering can, walk across a room and place it on a lower shelf. The same model checkpoint was also used on other robot platforms, including Franka Duo, highlighting its ability to work across different robotic embodiments. Google said the model also improves dexterity. It can operate Apollo 2's five-fingered robotic hands to perform tasks such as tying trash bags, sealing Ziploc bags, and unscrewing light bulbs. On industrial platforms using two-finger grippers, it can perform precise insertion tasks, tool handling, and object placement. The company is also making its embodied reasoning model, Gemini Robotics ER 2, available through Google AI Studio and in private preview on Gemini Enterprise Agent Platform. The vision-language-action and on-device models are being released to early-access partners. Google says Gemini Robotics ER 2 serves as the robot's high-level reasoning engine, allowing it to understand spoken instructions, break them into multiple steps, and monitor progress until a task is complete. The updated model can execute task sequences lasting several minutes while making hundreds of decisions. It can also recover from failed actions and adapt to changing environments instead of following fixed routines. Another new capability is multi-robot collaboration. Different types of robots can now communicate and divide work to complete jobs that would be difficult for a single machine to perform alone. For applications where internet connectivity is unavailable, Google introduced Gemini Robotics On-Device 2. The company said the model runs locally on robotic hardware and can adapt to entirely new dual-arm robot designs with fewer than 200 training examples collected over just a few hours. Google also introduced ASIMOV-Agentic, a new benchmark designed to evaluate robotic safety, including whether robots refuse unsafe actions, recognize uncertainty, and request human assistance when needed. The company said Gemini Robotics ER 2 also improves human proximity detection, allowing robots to stop safely when people move too close.

Google updates its AI software for robots

Why it matters: It's part of a growing trend toward using AI models to make humanoid and other robots more versatile. Driving the news: The effort includes both models that control robots' individual physical actions and those designed to handle higher-level task planning. * The update builds on the push Google announced last year to bring its Gemini models to robots. What they're saying: "Gemini Robotics 2 enables robots to reason through every movement, unlocking a broad range of tasks," Google said in a blog post. * "For example, it can enable a humanoid to walk, crouch, stretch, and manipulate objects to clean up a cluttered room. It can even team up with other robots to finish the job faster." The big picture: Google's move comes as US companies may have a tougher time getting their hands on the robots needed to run such software.

Gemini Robotics 2 brings whole body intelligence to robots

From feet to fingertips -- we are teaching robots intelligent whole-body control, fine dexterity, and teamwork to complete a broad range of complex tasks For decades, we've dreamed of robots that can seamlessly step into our world and lend a hand. Now, that vision takes a significant stride forward. Most robots are pre-programmed or teleoperated for narrow, repetitive task sequences. They lack the ability to truly learn for themselves or adapt to unpredictable environments. Moreover, transferring learned skills from one robot body to another remains incredibly difficult. To take on the hardest problems at scale, robots of every shape and size need AI models giving them the ability to think, act, and interact intelligently to safely complete tasks. We demonstrated how Gemini's multimodal understanding could drive real-world action with Gemini Robotics. Today, we are introducing Gemini Robotics 2 - the intelligence layer powering the next generation of truly adaptable robots. As it takes its first literal steps, this major advance unlocks intelligent whole-body control, advanced dexterity, and multi-robot collaboration. Gemini Robotics 2 enables robots to reason through every movement, unlocking a broad range of tasks. For example, it can enable a humanoid to walk, crouch, stretch, and manipulate objects to clean up a cluttered room. It can even team up with other robots to finish the job faster. And this profound intelligence can also run locally on-device while seamlessly adapting to entirely new robotic bodies in just a few hours. We are making this possible through three highly capable models:

Introducing Gemini Robotics ER 2

Summaries were generated by Google AI. Generative AI is experimental. For robots to assist humans in everyday environments, accurate spatial reasoning is not enough. Robots must also think fast, timing their decisions and reasoning with the real-time speed of the physical world. That's why today we're launching Gemini Robotics ER 2, our most capable "embodied reasoning" model for robotics. Think of Gemini Robotics ER 2 as a high-level brain for robots. It allows robots to chat with humans, understand the physical world, and plan multi-step tasks. It then hands off motor execution to any given lower level vision-language-action (VLA) model. Gemini Robotics ER 2 can also natively call tools like Google Search to find information, or any other user-defined function. The design of Gemini Robotics ER 2 allows the robot to "think" about what comes next while simultaneously performing its actions. Gemini Robotics ER 2 represents a significant upgrade over Gemini Robotics ER 1.6. By watching continuous video feeds, robots can now track their own progress, adapt if something goes wrong, and know exactly when to move on to the next step. We are also introducing multi-robot collaboration, enabling robots to work together in shared spaces and complete complex workflows a single robot could not do alone. Gemini Robotics ER 2 is now publicly available to developers via the Gemini API, Google AI Studio, and in private preview on Gemini Enterprise Agent Platform. To help you get started, we're sharing examples of how to configure the model and prompt it to power more useful physical AI tasks. Advancing physical agentic capabilities Most tasks in the physical world are complex and require multiple steps to complete. Gemini Robotics ER 2 is a physical agent, orchestrating steps for the robot and enabling it to self-correct, and generalize to more novel situations. To build an agentic setup, developers can declare low-level control interfaces -- like Vision-Language-Action (VLA) models or navigation APIs -- as tools, and stream multimodal video, audio, or text directly into the model. Gemini Robotics ER 2 improves this tool orchestration workflow. We can evaluate its performance with robots in simulation, using real-world robot control, and even pair it with a human controlling the robot remotely.

Google Introduces Gemini Robotics Er 2

Google introduced Gemini Robotics ER 2 represents a step change in powering robots with video understanding, task orchestration, and multi-robot collaboration ? making it possible for robots to be more helpful in the physical world. For robots to assist humans in everyday environments, accurate spatial reasoning is not enough. Robots must also think fast, timing their decisions and reasoning with the real-time speed of the physical world. That?s why today we?re launching Gemini Robotics ER 2, our most capable ?embodied reasoning? model for robotics. Think of Gemini Robotics ER 2 as a high-level brain for robots. It allows robots to chat with humans, understand the physical world, and plan multi-step tasks. It then hands off motor execution to any given lower level vision-language-action (VLA) model. Gemini Robotics ER 2 can also natively call tools like Google Search to find information, or any other user-defined function. The design of Gemini Robotics ER 2 allows the robot to ?think? about what comes next while simultaneously performing its actions. Gemini Robotics ER 2 represents a significant upgrade over Gemini Robotics ER 1.6. By watching continuous video feeds, robots can now track their own progress, adapt if something goes wrong, and know exactly when to move on to the next step. We are also introducing multi-robot collaboration, enabling robots to work together in shared spaces and complete complex workflows a single robot could not do alone. Gemini Robotics ER 2 is now publicly available to developers via the Gemini API, Google AI Studio, and in private preview on Gemini Enterprise Agent Platform. To help you get started, we?re sharing examples of how to configure the model and prompt it to power more useful physical AI tasks. Advancing physical agentic capabilities Most tasks in the physical world are complex and require multiple steps to complete. Gemini Robotics ER 2 is a physical agent, orchestrating steps for the robot and enabling it to self-correct, and generalize to more novel situations. To build an agentic setup, developers can declare low-level control interfaces ? like Vision-Language-Action (VLA) models or navigation APIs ? as tools, and stream multimodal video, audio, or text directly into the model. Gemini Robotics ER 2 improves this tool orchestration workflow. We can evaluate its performance with robots in simulation, using real-world robot control, and even pair it with a human controlling the robot remotely. In robotics, high-level reasoning depends on execution speed. Gemini Robotics ER 2 integrates into the Gemini Live API, using a bidirectional streaming endpoint optimized for latency-sensitive tasks. The result is fluid orchestration: Gemini Robotics ER 2 commands action models and robotics APIs to complete multi-step tasks without the jarring ?stop-and-think? pauses. To illustrate this, we?ve built a demo with Spot from our partners at Boston Dynamics. We use Gemini Robotics ER 2 to orchestrate Spot APIs, such as navigation and manipulator movement, creating an interactive robot that fetches objects for you. The code is available on Github with other examples. Unlocking temporal intelligence for robust task completion One of robotics? hardest challenges is knowing when a task is done. Gemini Robotics ER 2 brings a step-change in video understanding and progress tracking to verify that complex tasks ? such as tightening a light bulb or tying a trash bag ? are complete to specification before switching to the next task. In this update, we?ve made progress on two foundational capabilities for task progress understanding: progress classification and moment finding. Continuous progress classification Progress classification refers to a robot?s ability to track progress towards task completion. In our evaluations, we assign each frame in a video feed into five levels of progress (0-20%, 20-40%, 40-60%, 60-80%, 80-100%). By quantifying task progress, Gemini Robotics ER 2 provides robots with real-time situational awareness, and allows them to adjust actions on the fly or retry failed steps without restarting an entire workflow. Precision moment-finding Moment-finding measures a model's ability to identify the exact video frame where a critical event takes place (i.e. when to stop pouring coffee into a cup). Gemini Robotics ER 2 achieves significant gains in performance on moment finding, enabling robots to precisely switch between tasks, verify success and suggest corrections. Multi-robot collaboration No single robot fits every task ? a wheeled rover excels indoors, while a humanoid robot may excel at uneven terrain. Gemini Robotics 2 enables multi-robot collaboration, allowing diverse machines to communicate via a shared semantic understanding to handoff and complete complex tasks. See how Gemini Robotics ER 2 enables Apptronik?s Apollo 2 and Franka F3 Duo to collaborate here. Improving general spatial intelligence Gemini Robotics ER 2 advances our core spatial reasoning capability, as measured by three benchmarks: Success/failure detection: Now operates on raw video feeds rather than static snapshots to catch mid-execution failures like spills, slips, or misalignments. General instrument reading: Extends beyond circular dials and sight glasses to include digital displays, linear scales, rulers, and liquid thermometers. We tested it across 10 different types of instruments. Enhanced spatial VQA: Improves Visual Question Answering throughGemini?s advancements in multi-modal understanding. Advancing safety for embodied intelligence Gemini Robotics ER 2 is our safest model, achieving significant gains on Safety Instruction Following and Human Proximity benchmarks, which evaluate how a model adheres to physical constraints during reasoning tasks and spatial awareness for detecting humans. We found that Gemini Robotics ER 2 successfully halts a humanoid robot when a person is nearby and autonomously resumes work only once the area is clear. To advance safety for physical agents, we?re introducing a benchmark that evaluates a foundation model's ability to act as a safe VLA orchestrator by testing its capacity to enforce safety constraints, monitor the environment, assess physical feasibility, and seek human clarification. For details, see our safety technical report. Looking ahead, our plans are to push these models towards even more complex tasks to accelerate the development of helpful robots and support the robotics community.