DeepMind teaches robots how to read gauges with Gemini Robotics-ER 1.6

Getting a robot to pick up a plastic cup is old news, but getting it to inspect a chemical plant and accurately read a physical needle gauge without hallucinating is where things get tricky.

Google DeepMind just launched Gemini Robotics-ER 1.6, an upgraded embodied reasoning model designed to act as the high-level brain for physical machines.

The model outperforms both its predecessor Gemini Robotics-ER 1.5 and the generic Gemini 3.0 Flash, sharpening core spatial capabilities like pointing, counting, and multi-view task success detection.

Developers can already access the model through the Gemini API and Google AI Studio, complete with a starter Colab notebook.

Seeing from multiple angles and calling search

Spatial reasoning in Gemini Robotics-ER 1.6 relies heavily on pointing as an intermediate reasoning step.

The model can point to identify optimal grasp points, calculate movement trajectories, or follow complex constraint prompts, like pointing only to items small enough to fit inside a designated cup.

Autonomy in robotics often breaks down because machines do not know when a task actually succeeded. With most modern rigs carrying both an overhead camera and a wrist camera, synthesizing multiple video feeds is brutal.

DeepMind boosted the model's multi-view reasoning across dynamic environments and bad lighting, giving robots the judgment to decide whether an attempt failed and needs a retry, or if it is safe to proceed.

It can also natively call tools like Google Search to pull context, or dispatch sub-tasks to downstream vision-language-action models and user-defined functions.

The agentic vision behind reading analog gauges

The marquee real-world showcase comes from DeepMind's partnership with Boston Dynamics.

Industrial facilities rely on quadruped robots like Spot to patrol facilities packed with legacy analog gauges, thermometers, and chemical sight glasses.

Reading these is a nightmare due to reflection, perspective distortion, and tiny tick marks. DeepMind tackled this using agentic vision, pairing visual reasoning with code execution.

The model zooms in on the gauge, drops spatial points, writes and runs code to calculate ratios and intervals, and interprets the text units. According to Future Tools, this workflow hits a 93% accuracy rate on gauges and sight glasses.

DeepMind also tightened safety constraints. The model refuses tasks violating physical limits, like handling liquids or lifting over 20 kilograms, and improves hazard detection on real-world injury reports by 6% in text and 10% in video over Gemini 3.0 Flash.

Why it matters

We think DeepMind's architectural division here is spot on. Rather than trying to force one monolithic model to handle both high-level common sense and millisecond-level joint motor controls, it focuses purely on being an executive reasoning engine.

For enterprise developers and robotics startups, this is immediately practical. Retrofitting industrial plants usually stalls because legacy dial meters require custom, fragile computer vision pipelines. An API that reads analog instruments off-the-shelf drastically lowers the barrier to deploying inspection robots.

Using code execution to calculate gauge fractions is an exceptionally clever way to avoid visual hallucination. The catch, however, is reliability in the field. When physical inspection relies on cloud API round-trips and tool-calling chains, spotty connectivity in heavy industrial facilities could quickly become an operational headache.