The robot shows up every day. It walks past the same pressure gauges, peers at the same sight glasses, checks whether anything has pooled where it should not. It does not get tired. It does not miss Tuesday.
That is the pitch for putting a robot in an industrial facility. The question has always been whether the robot is good enough to act on what it sees — or whether it cries wolf often enough that the operators stop listening.
Boston Dynamics has an answer: 80 percent.
In an interview with IEEE Spectrum this week, Marco da Silva, vice president and general manager of Spot, said that is the threshold where Spot becomes useful rather than annoying. Below 80 percent accuracy, operators stop trusting the robot's readings and start ignoring them. Above it, the system is reliable enough to act on. The company is now rolling out an upgrade that it believes clears that line.
The upgrade is Gemini Robotics-ER 1.6, a reasoning model from Google DeepMind, integrated into Spot's inspection software. DeepMind released the model publicly on April 14th, with Boston Dynamics as the first commercial partner. Until now, most inspection robots followed pre-programmed routes without understanding what they saw. Gemini Robotics-ER 1.6 is what the AI industry calls a vision-language-action model — it looks at the robot's cameras, understands the context, and decides what to do next. It can read a pressure gauge, understand what the reading means, and flag whether something is wrong.
The key technical advance is called agentic vision. When Spot points its camera at a gauge, the model does not just read the number — it zooms into the image, traces the needle's position against the tick marks, runs code to calculate the precise value, and cross-references it against what it knows about that type of instrument. It does this for analog gauges with needles, digital readouts, and vertical sight glasses where liquid levels need to be estimated accounting for camera distortion. The result, DeepMind claims in its own benchmarks, is accuracy down to sub-tick — finer than what a human inspector can reliably read by eye.
Da Silva said the company is rolling the new capabilities out through beta programs with a smaller set of customers before broadly advertising features they are confident will work, according to Boston Dynamics' own blog post. Several thousand Spot robots are already commercially deployed across power plants, offshore oil rigs, nuclear decommissioning sites, and manufacturing facilities, according to IEEE Spectrum. Those deployments give Boston Dynamics a rare advantage in the robotics industry: a large, real-world feedback loop.
Carolina Parada, head of robotics at Google DeepMind, told IEEE Spectrum that the decision to build vision-only models — no touch sensors, no force feedback — was a deliberate bet on what data is available. There is abundant visual data on the internet for training models to understand objects and environments. There is very little data combining touch and vision together. "If we had enough data with touch information, we could easily learn it, but there is not a lot of data with touch sensing on the internet," she said. The inspection deployments provide that real-world visual data.
There is a data exchange in the arrangement. Customers who use the new Gemini-powered inspection capabilities on their Spot robots are required to share their inspection footage back to Boston Dynamics, according to IEEE Spectrum. The company says this helps improve the models. It also means every gauge read, every hazard spotted, every facility walk-through is feeding Google's robotics research.
The vision-only approach has limits. Poor lighting, heavy occlusion, and unusual gauge configurations are known failure modes for systems that rely purely on cameras. The 80 percent threshold — a number Boston Dynamics gave, not an independent measurement — refers to something the company is confident it is hitting in beta. Hard environments, like a power plant with steam and confined spaces, may not look like the beta sites.
The robot that walks the facility every day is not yet a replacement for a human inspector. It is closer to a very attentive junior technician who never sleeps and never misses a scheduled walk-through. Whether that is enough depends entirely on what happens when it cries wolf — and whether, this time, the wolf is actually there.
