When Spot reads a pressure gauge wrong, a human decides whether to evacuate. When it reads wrong at 79 percent accuracy, they don't decide — they ignore it.
Boston Dynamics is retrofitting its robot dogs with better eyes and a data bill. The upgrade, powered by Google DeepMind's Gemini Robotics-ER 1.6, lets Spot read analog gauges, pressure valves, and thermometers — the instruments that tell a technician whether a machine is about to fail or already on fire. In internal benchmarks, accuracy on instrument-reading tasks jumped from 23 percent with the previous model to 98 percent, according to IEEE Spectrum. That is the jump where the robot stops being a novelty and starts being useful.
But there is a cost that does not appear in the demo video.
Customers who want the upgrade must enroll in Boston Dynamics' AIVI-Learning program, and AIVI-Learning comes with a data-sharing requirement. The model that reads your facility's gauges is trained, in part, on your facility's gauges. "We require data sharing to train and improve the models that power AIVI-Learning, ensuring they contextually understand the specialized, complex use cases unique to your site," Boston Dynamics explains on its blog. The company confirmed the transition went live for enrolled customers as of April 8, 2026.
Spot operators should not expect the robot to hit that 98 percent accuracy mark on day one. Marco da Silva, vice president and general manager of Spot at Boston Dynamics, told IEEE Spectrum that the useful threshold sits just north of 80 percent accuracy. Below that, operators start ignoring the alerts. Above it, they actually respond. Getting from 80 to 98 requires the model to see enough of your specific environment, your gauges, your failure modes. That learning happens on your data.
Boston Dynamics has several thousand Spots operating across customer sites, according to IEEE Spectrum, one of the few companies actually deploying legged robots at scale. Every patrol every robot runs is a training signal. The upgrade that makes Spot more useful to a facility operator also lets Google and Boston Dynamics accumulate a proprietary dataset of how industrial equipment actually behaves across thousands of different sites. The gauge-reading is the bait. The data is the catch.
Google DeepMind has made Gemini Robotics-ER 1.6 available to developers via the Gemini API and Google AI Studio starting April 14, 2026, according to the company's blog. The safety claims are Google's own. The model demonstrated superior compliance with Gemini safety policies on adversarial spatial reasoning tasks, per the same post, and the accuracy benchmarks are internal, not independently tested.
Carolina Parada, head of robotics at Google DeepMind, told Ars Technica that the baseline model without the agentic vision system still achieves 86 percent accuracy, compared to 67 percent for Gemini 3.0 Flash. The jump from 67 to 86 to 98 is real. Whether it is 98 in a lab and 82 in a steam-fogged power plant is a different question, and the data-sharing requirement means that is a question Boston Dynamics will answer better than anyone else.
