Google DeepMind unveils Gemini Robotics-ER 1.6 for robot autonomy and instrument reading

On April 14, 2026, Google DeepMind introduced Gemini Robotics-ER 1.6, an update to its reasoning model that acts as a high-level cognitive layer for robots in the physical world. The point of version 1.6 is not to add yet another VLA, but to give a robot more precise spatial reasoning: the model understands a scene better, counts objects, determines whether a task has been completed and, for the first time, reliably reads complex instruments such as pressure gauges, level gauges and digital displays.

DeepMind calls Gemini Robotics-ER a reasoning-first model for embodied AI. It is meant for situations where a robot needs more than object recognition: it has to understand relationships between items, choose a grasp point, check constraints and decide what to do next. Version 1.6 notably strengthens pointing, the ability to indicate objects and use those points as an intermediate step in reasoning. This helps the model count items more accurately, compare sizes, plan trajectories and follow instructions with spatial conditions.

The model also gets improved multi-view understanding: it does a better job of assembling a single picture from several cameras, for example an overhead one and one mounted on the manipulator. For developers, Gemini Robotics-ER 1.6 is already available through the Gemini API and Google AI Studio, along with Colab examples for setup and prompt engineering.

The most notable new skill is instrument reading, which grew out of joint work with Boston Dynamics. In industrial facilities, robots regularly encounter thermometers, circular pressure gauges, sight glasses and vertical level indicators, where what matters is not simple image classification but accurate extraction of the value. To do this, Gemini Robotics-ER 1.6 uses agentic vision, which combines visual reasoning with code execution. The model first zooms in on the relevant region, then marks key points, estimates intervals and proportions, and finally matches the result against the scale, units of measurement and context.

According to Google DeepMind, on the instrument reading task version 1.5 scored 23%, Gemini 3.0 Flash 67%, Robotics-ER 1.6 itself 86%, and with agentic vision 93%. This is no longer a demonstration that a robot "sees" a gauge, but a step toward a scenario in which it makes its own rounds, reads the values and understands what they mean.

Another important area is success detection and safety. For an autonomous robot, it is not enough to be able to start a task; it must know when the task is truly complete and when the attempt needs to be repeated. DeepMind says the model handles success detection better even in dynamic scenes, with partial occlusions and ambiguous viewing angles.

Compliance with physical constraints has also improved: for example, the system should now better respect prohibitions such as "do not pick up liquids" or "do not lift objects heavier than 20 kg." In tests on recognizing hazardous situations from text and video, the Gemini Robotics-ER family improved on Gemini 3.0 Flash by 6% and 10%, respectively.

Google itself separately notes a limitation: the model is not intended for safety-critical uses such as medicine, transportation and other environments where an error could lead to injury or damage.

The practical meaning of the release is that Google is gradually turning embodied reasoning from a research topic into an infrastructure layer for robotics. Gemini Robotics-ER 1.6 does not control the hardware directly, but gives robots a stronger high-level reasoning layer that can be linked with VLA models, search and external functions. For the industry, it is also a signal that the interface between language models and robots is moving closer to commercial use. If this combination proves itself outside the lab, the market will get robots that do not just follow a script, but interpret their surroundings, verify results and read real instruments without a human in the loop.