Google DeepMind has introduced Gemini Robotics ER 2, a new embodied reasoning model designed to act as a high-level decision-maker for robots. The pitch is simple but ambitious: instead of merely responding to a command or recognizing an object in a snapshot, a robot should be able to watch an unfolding scene, understand its progress through a task, recover from mistakes and coordinate with other machines when the work calls for it.

That matters because the gap between a slick robotics demo and a genuinely useful robot is usually not basic movement. A robot may be able to grasp an item, open a drawer or navigate a hallway. The harder question is whether it can tell that the item slipped, the drawer did not fully close, coffee is about to overflow or a person has stepped into its workspace. Google is positioning ER 2 as the layer that handles that kind of ongoing judgment.

Gemini Robotics ER 2 is not, by itself, the model that moves a robot’s joints. It sits above lower-level vision-language-action, or VLA, systems that translate instructions and visual input into motor commands. Google describes ER 2 as the “brain” that can converse with people, interpret the physical environment, break down multi-step jobs and hand specific actions to whichever control model or robotics API is best suited to perform them.

In practice, that could mean a robot hears, “Bring me a snack from the kitchen,” and does more than locate a packet and attempt a grab. It needs to plan a route, identify the item, notice whether it has actually picked it up, deal with an obstruction, understand when the handoff has succeeded and then move on. That is the kind of orchestration Google says ER 2 is built for – with the system able to keep reasoning while the robot is already acting, rather than pausing between each instruction.

The timing is notable. Robotics has had a wave of progress in recent years, particularly around large AI models that connect what a machine sees with what it can do. But useful physical intelligence requires more than visual recognition and natural-language chat. It needs an understanding of space, sequence, timing and uncertainty – all while operating in environments that are far messier than a carefully prepared lab bench.

A robot that watches its work

The standout feature in ER 2 is continuous video understanding. Rather than treating a task as a collection of still images, Google says the model can monitor a live video feed to estimate how far along the robot is, detect when something goes wrong and decide when it is safe to advance to the next stage.

That sounds like a subtle upgrade, but it gets at one of robotics’ most stubborn problems: knowing when a task is actually complete. Consider tightening a light bulb. A robot needs to distinguish between touching the bulb, turning it partially, finishing the action and over-tightening it. Or consider tying a trash bag: it is not enough to make movements that resemble a knot. The machine needs to verify that the bag is closed to the required standard.

Google measures this through what it calls progress classification, where each video frame is placed into one of five task-completion ranges. On its evaluation, Gemini Robotics ER 2 reached 57.4% accuracy, outperforming the models Google compared it with. The company also reports 91.3% accuracy on “moment finding” tasks – identifying the precise moment a critical event occurs, such as when to stop pouring coffee – with a mean absolute timing distance of 0.96 seconds.

Those figures should be read as vendor-reported benchmark results rather than a guarantee of performance in every home, warehouse or factory. Still, they point to a useful shift in focus. Robots do not just need to perform actions; they need to observe the consequences of those actions in real time.

Google says ER 2 runs through the bidirectional Gemini Live API, which is intended for lower-latency streaming interactions. In one demonstration, the company paired the model with Boston Dynamics’ Spot robot and used it to orchestrate navigation and manipulator APIs so the robot could retrieve an object after a natural-language request.

Why multi-robot work matters

ER 2 also introduces multi-robot collaboration, a feature that could be more important than it initially sounds. Robotics companies often frame the future around a single general-purpose humanoid, but real workplaces rarely work that way. Different machines are good at different jobs: a wheeled robot may handle indoor transport efficiently, while a bipedal humanoid may be better suited to uneven terrain, stairs or tasks built around human-sized spaces.

Google’s approach is to give multiple robots a shared semantic understanding of the work. That means one machine can handle an early stage, communicate the state of the task and hand off to another robot with a different body, tool set or mobility profile. Google has highlighted a collaboration involving Apptronik’s Apollo 2 humanoid and Franka Robotics’ dual-arm F3 Duo system.

This is a more practical vision of automation than the idea of one robot that can do everything. In a warehouse, hospital or industrial site, a coordinated group of specialized machines may prove more effective than waiting for a single humanoid to become flawless at every task. The challenge, of course, is that coordination introduces new failure modes: robots must share an accurate view of the environment, avoid conflicts and understand who is responsible for what.

Google’s wider Gemini Robotics 2 release reflects that division of labor. Alongside ER 2, the company has introduced Gemini Robotics 2, a vision-language-action model for full-body robotic control, and Gemini Robotics On-Device 2, a more efficient VLA model intended to run locally on robots and adapt to new embodiments with only a few hours of data.

That split is worth paying attention to. Robotics is unlikely to be solved by one giant model doing everything. The emerging architecture looks more like a stack: a high-level planner, specialized action models, navigation and manipulation tools, sensors, safety systems and, increasingly, a way for all of them to communicate without creating delays or confusion.

Safety remains the hard part

Google is also making a case for ER 2 as its safest embodied reasoning model so far. The company says it has improved results on safety instruction following and human proximity benchmarks, including tests where a humanoid should halt when a person is nearby and resume only after the space is clear.

It is an important claim, but it also highlights why robotics safety cannot be reduced to a benchmark score. A language model can be excellent at describing a safe plan and still be part of an unsafe real-world system if its sensors fail, its control stack has unexpected behavior or the environment changes faster than the model can respond.

Google acknowledges part of that reality by focusing on ER 2 as an orchestrator, rather than the sole source of physical control. Developers can define lower-level action models and APIs as tools, while streaming video, audio and text to ER 2. The model can also call Google Search or developer-defined functions when an outside information source is needed.

That tool-calling flexibility could make robots more capable, but it creates another engineering responsibility: developers will need strict permissions, reliable guardrails and fallbacks for cases where the system is uncertain. A robot operating around people should not turn ambiguity into action.

The earlier Gemini Robotics ER 1.6 release showed the direction Google was already taking, with a focus on physical constraints, spatial reasoning, instrument reading and hazard recognition from text and video. ER 2 extends that work by emphasizing continuous progress tracking, faster orchestration and team-based robotics.

Available to developers now

Google says Gemini Robotics ER 2 is publicly available through the Gemini API and Google AI Studio, while access through the Gemini Enterprise Agent Platform is in private preview. The company has also published sample code for developers looking to connect the model with robotic tools and workflows.

For developers, the immediate story is less about buying a ready-made household robot and more about experimenting with a new planning layer for physical AI. ER 2 may be useful wherever a robot must manage long, variable tasks that require observation and correction: logistics, industrial inspection, lab work, retail backrooms and eventually some forms of assistance in public or home settings.

The bigger takeaway is that Google is pushing robotics beyond the familiar prompt-response model of generative AI. A robot in the real world cannot simply produce a plausible answer and move on. It must keep watching, decide when an attempt has failed, understand what changed and choose the next action without losing track of the larger job.

That is a high bar, and the field still has a long way to go before robots handle everyday environments with the reliability people expect. But Gemini Robotics ER 2 makes Google’s strategy clear: the next stage of physical AI will not be defined only by better hands, faster legs or more realistic humanoid designs. It will depend just as much on whether robots can reason through time – and recognize the difference between starting a task, appearing to finish it and actually getting it right.