DeepMind built a robot brain that can read analog gauges with precision finer than the smallest marked increment (sub-tick accuracy, in the jargon), and nobody asked for it.
Gemini Robotics-ER 1.6, the latest version of Google's embodied reasoning model, was designed to improve how robots understand physical spaces: how to navigate clutter, manipulate objects, reason about where things are in three dimensions. While testing it with Boston Dynamics on their Spot quadruped robot, researchers found it could read analog pressure gauges, dial indicators, and sight glasses with better precision than existing automated systems. It was not the point. It was not even on the roadmap.
"We were not planning to develop this capability," the DeepMind team wrote in their blog post. "It is a use case we discovered through close collaboration with our partner, Boston Dynamics."
The instrument reading capability works through a multi-step reasoning process the team calls agentic vision. When the model encounters a gauge, it first zooms into the image to resolve fine gradations that would blur at normal resolution. It then uses pointing gestures to estimate proportions and intervals, and executes code to compute the final reading. The result is accuracy DeepMind describes as sub-tick, meaning it can distinguish between markings on a gauge that are smaller than the interval between hash marks. For a standard dial gauge with 10-pound tick marks, sub-tick means estimating the needle position within a pound or less.
This matters because analog instrumentation is everywhere in industrial infrastructure, and it has resisted automation for decades. Power plants, water treatment facilities, oil and gas refineries, and manufacturing floors are full of pressure gauges, temperature dials, and level indicators that human inspectors walk past with clipboards. Vision systems trained on photographs of gauges struggle because lighting varies, angles distort apparent needle position, and reflections on glass covers introduce errors. The problem is unglamorous. Nobody writes grant proposals to automate gauge reading. It is, as DeepMind's blog framed it without apparent irony, "the unsexy problem."
But it is also a billion-dollar problem. Boston Dynamics, the robotics company behind Spot, has already deployed more than 1,000 inspection points across industrial facilities worldwide. Those deployments require routine inspection rounds, and every one of them involves a human reading gauges that machines cannot read reliably. Gemini Robotics-ER 1.6 changes that calculus. Spot already navigates those facilities autonomously. If the robot brain can now read the instruments too, the inspection loop closes without a human in the loop.
The timing matters. Gemini Robotics-ER 1.6 is available to developers via the Gemini API and Google AI Studio starting today. That puts the capability into any application that can call the API and connect to a camera-equipped robot. The barrier to deployment is lower than a full robotics system: any facility already running Spot or a comparable platform can add instrument reading by updating the model. The analog-to-digital conversion that industrial automation has struggled to solve for thirty years is now a software update.
The capability is a consequence of what makes Gemini Robotics-ER 1.6 different from its predecessors. The model shows significant improvement over both Gemini Robotics-ER 1.5 and Gemini 3.0 Flash specifically in spatial and physical reasoning. That is the core achievement: teaching a model to understand how objects exist in three-dimensional space and how forces, angles, and proportions relate to each other. Reading a gauge is a spatial reasoning task. The needle's position encodes information about pressure or temperature, and that encoding depends on the dial's geometry, the angle of view, and the relationship between the tick marks and the labeled values. A model that reasons well about space reasons well about instruments.
DeepMind is not the only lab working on this problem. Physical reasoning has been an active research frontier for several years, and multiple groups have published results on robotic manipulation, gauge reading, and instrument interpretation. What distinguishes the DeepMind result is the combination of sub-tick accuracy and the agentic vision pipeline, the multi-step process of zooming, pointing, and code execution rather than a single-pass classification. Whether that pipeline is proprietary to Gemini Robotics-ER 1.6 or describes a general approach that other labs can replicate is not yet clear from the published work.
The more pressing question is what else is hidden in these models. DeepMind found instrument reading by accident, while testing something else. Gemini Robotics-ER 1.6 is a foundation model, a single model trained once that can then be adapted to many tasks. Foundation models are characterized by emergent capabilities: abilities that appear in the model without being explicitly trained, that the researchers did not design and did not predict. Reading analog gauges was not a training objective. It emerged because the model's spatial reasoning became good enough to solve the problem.
Every major lab has unpublished capability catalogs, a list of things their models can do that they have not yet announced. The history of AI progress is partly a history of accidental discoveries: capabilities that existed in a model for months before anyone noticed they were there, or that appeared first in research settings and only later revealed their commercial significance. DeepMind stumbled into instrument reading. The question is what else is sitting undiscovered in frontier models, waiting for someone to test the right task with the right robot in the right facility.
Boston Dynamics has been commercially deploying Spot since September 2019. The company has had five years to identify where robots add the most value in industrial inspection. Gauge reading is high on that list. Gemini Robotics-ER 1.6 is available now. The inspection loop is closing.
What to watch: whether Google positions this as a vertical product for industrial inspection or keeps it as a general API capability. If the former, it is a direct challenge to the existing inspection software market, companies selling gauge reading as a standalone service may find their core differentiator is now a baseline feature in a robot brain. If the latter, it becomes infrastructure: invisible, standardized, and cheap, with the value flowing to whoever integrates it best.
