Gemini Robotics-ER snaps to 93%
- Google’s Gemini Robotics‑ER 1.6 raised industrial gauge‑reading accuracy and is being used on mobile robots. (x.com) - The model hit about 93% accuracy on industrial gauges, a roughly 70‑point jump over the prior 1.5 version. (x.com) - That performance gain is already showing up in field deployments, including integrations on Boston Dynamics' Spot robot. (x.com)
Robots that inspect factories are getting better at one of the dullest jobs on the floor: reading gauges. Google DeepMind said its Gemini Robotics-ER 1.6 model reached 93% accuracy on industrial instrument reading. (deepmind.google) Industrial inspections often mean checking analog dials, sight glasses, thermometers, and digital displays to see whether equipment is running inside safe limits. DeepMind said the new model can read those instruments through a robot’s cameras and use that information in its next action. (deepmind.google) Google released Gemini Robotics-ER 1.6 on April 15, 2026, and said the instrument-reading feature was developed with Boston Dynamics. The company made the model available to developers through the Gemini Application Programming Interface and Google AI Studio. (deepmind.google) DeepMind said the 93% score came from “agentic vision,” which means the model can direct the robot to gather better views before deciding what it sees. In the company’s comparison, Gemini Robotics-ER 1.5 scored 23% on the same instrument-reading task, while Gemini 3.0 Flash scored 67%. (deepmind.google) Boston Dynamics is already tying that software into Spot, its four-legged inspection robot used at industrial sites. The company says Spot is built to automate sensing and inspection and is already used to collect visual, thermal, and acoustic data in facilities. (bostondynamics.com, bostondynamics.com) In a separate post last week, Boston Dynamics said it had integrated Gemini and Gemini Robotics-ER 1.6 into Orbit AIVI-Learning, its inspection software stack. The company said the update lets Spot and Orbit “continuously learn” about a facility and deliver a more capable artificial intelligence experience for inspection work. (bostondynamics.com) Boston Dynamics has been moving toward this use case for months. In another April post, it showed Spot using Gemini Robotics-ER 1.5 for “embodied reasoning” in a home setting and said the demo grew out of a 2025 internal hackathon. (bostondynamics.com) The practical appeal is straightforward: many plants still rely on workers to walk routes, look at instruments, and log readings by hand. Boston Dynamics markets Spot as a way to keep people out of hazardous areas while gathering more frequent inspection data. (bostondynamics.com, bostondynamics.com) DeepMind also said 1.6 improved other robotics reasoning tasks, including pointing, counting, understanding cluttered scenes, and checking whether a job was completed. The company described the model as its “safest robotics model to date” and said it showed gains on physical-safety evaluations. (deepmind.google) For factory operators, the near-term test is not whether a robot can walk a route, but whether it can read a dial correctly often enough to trust the alert. DeepMind and Boston Dynamics are now saying that number is high enough to put the feature into live inspection products. (deepmind.google, bostondynamics.com)
