Spot can now read analogue gauges with almost 98% accuracy.

New model brings real-world reading to mobile robots

Boston Dynamics’ four-legged robot Spot has taken a step closer to being a practical industrial inspector. The company has been running trials that pair the robot with Google DeepMind’s latest robotic AI, Gemini Robotics-ER 1.6, which was announced on 14 April 2026. The combination lets Spot not only move through factories and warehouses but also interpret analogue instruments — thermometers, pressure gauges and sight-glass indicators — with performance levels that were, until very recently, out of reach for roaming robots.

In many industrial settings, the real challenge goes beyond just recognizing images. It’s the messy, real-world detail: multiple needles, overlapping tick marks, reflections on glass, labels at odd angles and liquid menisci seen through sight glasses. Those elements make instrument reading a complex visual-reasoning task rather than a simple picture-match. And that’s exactly what Google DeepMind set out to solve.

How Gemini Robotics-ER 1.6 works

Gemini Robotics-ER 1.6 is described by Google DeepMind as a high-level reasoning model for robots — a system that doesn’t just see, it plans and acts. The model builds on a capability Google calls "agentic vision", which mixes visual reasoning with the ability to execute small pieces of code while the robot inspects an image. The visual reasoning can point at or highlight elements in a scene and then run logic to combine those observations into a final answer.

That combination really changes the game. Where previous models tended to guess based on a single frame, the new setup can parse multiple visual cues and check them against one another. It also supports multi-view reasoning, using several camera streams to build a fuller picture of a dial or a sight glass as the robot moves around it.

That helps where reflections, occlusions or awkward perspective would otherwise fool a single view.

The numbers speak for themselves. Google DeepMind’s tests show the older Robotics-ER 1.5 model managed about 23% accuracy on instrument-reading tasks. Gemini 3.0 Flash — which first introduced agentic vision in January 2026 — achieved 67% accuracy in comparable tests. The new Robotics-ER 1.6 with agentic vision reaches around 98% accuracy on those same tasks, the company says. Even without agentic vision, the base Robotics-ER 1.6 model can hit roughly 86% accuracy by pointing to visual elements and reasoning over them.

Why factories and plants care

Robotic inspection is an attractive proposition for operators of large industrial sites. Routine checks of pressure gauges, temperature monitors and liquid levels can be time-consuming and sometimes hazardous. Boston Dynamics has been exploring Spot as a mobile inspector that could relieve human workers of repetitive checks, and as a way to gather more frequent data from equipment spread across a site.

Boston Dynamics’ parent, Hyundai Motor Group, runs large factories where automated inspections could be useful. Boston Dynamics has said it wants to test both quadruped and humanoid robots in a range of facilities. Those trials are a natural fit for an AI that improves a robot’s ability to understand the physical world rather than simply recognise objects in a curated dataset.

Mastheads

The real value comes from handling various instruments without needing custom training for each one. Gemini Robotics-ER 1.6’s multi-view and agentic vision techniques are geared towards that generality: the model is intended to generalise across instrument types, angles and lighting conditions rather than requiring a tailored model per gauge.

Technical limits and remaining challenges

The high accuracy seen in tests is encouraging. But real-world deployments will still present edge cases. Instruments behind grime or heavy condensation, gauges with damaged needles, or dials with non-standard markings will test robustness. In addition, industrial sites present logistical hurdles: Wi‑Fi dead zones, variable lighting and moving machinery can make the simple act of getting a clear camera view.

Robots have to figure out when to take action and when to alert a human. That’s partly a software problem and partly an operational one. A model that can flag uncertain readings and either retake a measurement from a different angle or hand the situation to a human inspector is more valuable than one that simply returns a best-guess value with no confidence estimate.

By combining visual pointers, multiple camera views, and simple code execution, Gemini Robotics-ER 1.6 equips robots to handle these checks. But deployment will demand integration with site safety rules, maintenance workflows and existing control systems. Boston Dynamics and its industrial partners will need to stitch the robot’s outputs into a larger maintenance and monitoring pipeline.

Where the technology could go next

This progress marks just one step in a longer journey. Robots that can perceive and reason about physical instruments unlock more complex inspection tasks: looking for signs of corrosion, estimating liquid levels behind opaque surfaces using small visual cues, or monitoring fluid flow through sight glasses. Each of those tasks asks for a different kind of visual reasoning and different thresholds for confidence.

Boston Dynamics is working on both four-legged and humanoid robots because they expect these machines to work in many environments—four-legged ones handle rough terrain, while humanoids can reach panels and flip switches. Google DeepMind’s role is to give those devices a more capable "brain" for reasoning about objects and actions in space, not just a better camera or classifier.

There are also regulatory and workforce questions around the wider adoption of robotic inspectors. How are readings from robots certified? Who signs off on maintenance actions suggested by an automated system? Those are organisational issues that will influence how quickly companies adopt this kind of tech, and what limits they place on the robot’s autonomy.

Early trials and industry reception

Boston Dynamics has already been trialling Spot in a range of industrial facilities. The trials aim to explore practical inspection routes, durability and how the robot integrates with human teams. Hyundai Motor Group’s factories are a logical place to run such tests given their scale and the parent company’s interest in robotics.

Industry observers say the ability to automate frequent, low-complexity checks could free technicians for higher-value work and reduce the time between anomaly detection and intervention. But they also say that trust in automated readings will build slowly — after repeated, verifiable success in live conditions. That incremental approach matches how many industrial operators adopt new monitoring technologies.

The technical gains here are measurable and specific. Google DeepMind reports big jumps in instrument-reading accuracy between model versions, and Boston Dynamics’ trials supply the mobility and durability a real site needs. The result is a practical route towards robotic inspection that’s more than a research demo.

Still, the jump from trials to routine use will be judged on day-to-day reliability, safety integration and how smoothly robots can fit into existing maintenance practices. For now, the headline is clear: an AI upgrade has turned a mobile robot into a much better instrument reader. The next stage will show whether that headline becomes a standard industrial tool.

Related Articles

Google DeepMind announced Gemini Robotics-ER 1.6 on 14 April 2026 and said the model raises instrument-reading accuracy from 23% to about 98% with agentic vision.

This article was created with AI assistance.