Google DeepMind updates Gemini model to help robots read gauges

In short: Google DeepMind announced a new version of its Gemini Robotics model that helps robots read factory instruments like analog gauges and thermometers.

What happened

Google DeepMind announced Gemini Robotics-ER 1.6 on April 14. The company says it is a “reasoning model for a robot,” meaning software that helps a robot plan what to do and then carry it out.

One clear example is inspection work in factories and warehouses. Robots like Boston Dynamics’ four-legged Spot can roam around and check analog thermometers and pressure gauges, the kind with needles and tick marks. DeepMind says the model also helps with visual inspections using sight glasses, which are small windows that let you see liquid levels inside tanks or pipes.

DeepMind links this progress to its collaboration with Boston Dynamics. Boston Dynamics has been testing Spot and other robots in industrial sites, including automotive factories connected to its owner, Hyundai Motor Group.

DeepMind says the model’s accuracy on instrument reading rose from 23% with the older Gemini Robotics-ER 1.5 to 98% with 1.6 when using a feature it calls “agentic vision.” In simple terms, that feature lets the system look at an image and take extra steps to work things out, like using a scratchpad while solving a math problem. DeepMind also says the model is better at using multiple camera views at once.

Why it matters

Factories rely on routine checks to catch problems early, but reading dials and liquid levels can be slow and easy to miss when people are tired. If robots can do these checks more reliably, it could reduce downtime and improve safety, but it also raises the stakes if a robot misreads something or moves unsafely around workers.

Source: Arstechnica