Gemini Robotics-ER 1.6: Powering real-world robotics tasks through enhanced embodied reasoning

Robots have gotten pretty good at picking things up and following scripted routines. Understanding what they're looking at, and knowing whether they actually finished the job, has been the harder problem. Google DeepMind's answer is Gemini Robotics-ER 1.6, the latest version of its embodied reasoning model, and it's available now through the Gemini API and Google AI Studio.

The headline feature is something Boston Dynamics pushed for: instrument reading. Industrial sites are full of analog gauges, thermometers, and chemical sight glasses that someone has to check by hand, or that a robot like Spot photographs on its rounds. Reading those accurately means dealing with needle positions, tick marks, camera distortion, and multiple decimal-place dials at once. Gemini Robotics-ER 1.6 handles this by zooming into an image, using code execution to estimate proportions between tick marks, and then applying general world knowledge to interpret the units. Marco da Silva, who runs the Spot program at Boston Dynamics, says this kind of reasoning is what lets Spot react to real-world conditions without a human in the loop.

Underneath that headline trick sits a more basic upgrade: pointing. DeepMind treats pointing as the building block of spatial reasoning for a robot brain, since it can express counting, grasp locations, motion paths, and constraints like