Google DeepMind's Gemini Robotics-ER 1.6 teaches robots to read gauges, lifting accuracy to 93% with Boston Dynamics' Spot as first showcase

Google DeepMind has given its robot reasoning model a skill that matters on factory and plant floors. On April 14, 2026 the Alphabet unit released Gemini Robotics-ER 1.6, a vision language model specialised in embodied reasoning, and made it available to developers the same day through the Gemini API and Google AI Studio under the model name gemini-robotics-er-1.6-preview. The headline addition is instrument reading, the ability to look at a pressure gauge, sight glass or thermometer and report the value.

DeepMind's own benchmark shows the size of the step. On its instrument reading evaluation, the previous Gemini Robotics-ER 1.5 succeeded 23% of the time and the general purpose Gemini 3.0 Flash 67%. ER 1.6 reached 86%, and 93% when allowed to use what DeepMind calls agentic vision. The company also reported better pointing, counting and success detection than ER 1.5, including from multiple camera views.

Agentic vision means the model takes intermediate steps instead of answering in one pass. According to the DeepMind blog post by Laura Graesser and Peng Xu, the model first zooms into the image to resolve small details on a dial, then uses pointing and code execution to estimate proportions and intervals before giving a reading. The instruments named include circular pressure gauges, chemical sight glasses, vertical level indicators, thermometers and digital readouts.

The model card, published on April 20, states that ER 1.6 is built on Gemini 3.0 Flash and accepts text, images, audio and video with a context window of 128,000 tokens, returning up to 64,000 tokens of text. It was trained on Google's Tensor Processing Units using Gemini 3.0 data plus additional embodied reasoning datasets. DeepMind says the model is not approved for safety critical uses such as healthcare or transportation, or settings where a malfunction could cause injury or property damage.

Boston Dynamics, a Hyundai Motor Group company, worked closely with DeepMind on the feature. Its Spot quadruped already walks routes through industrial sites capturing images of equipment, and ER 1.6 gives it a way to interpret what it photographs. Marco da Silva, a vice president at Boston Dynamics, said in DeepMind's post that instrument reading and more reliable task reasoning will let Spot see, understand and react to real world challenges completely autonomously.

The ER line plays a specific role in DeepMind's robotics stack. Gemini Robotics models output motor actions; the ER models do the thinking around them, deciding where objects are, what order to do things in and whether a task succeeded. Developers can call ER 1.6 from their own robot software and pair it with any controller, which makes it the most broadly available part of DeepMind's robotics offering.

Safety evaluation was a large part of the release. DeepMind tested ER 1.6 on version 2 of its ASIMOV benchmark of human centric safety scenarios and said it adheres to physical safety constraints substantially better than ER 1.5 in pointing and bounding box tasks. Against Gemini 3.0 Flash it reported improvements of 6% on text based safety scenarios and 10% on video scenarios.

The competitive field for robot reasoning models is crowded. NVIDIA offers its Cosmos Reason models and GR00T policies to robot makers. Physical Intelligence and Skild AI train end to end models that combine reasoning and action. What DeepMind brings is distribution through a commercial API that any developer can already use, and a model family tied to Gemini, Google's frontier model line.

For industrial operators the significance lies in a dull but costly job. Gauge rounds in refineries, chemical plants, power stations and data centres still rely heavily on technicians walking routes. A robot that can read analogue dials without retrofitting sensors removes one of the main barriers to automating inspections in older plants.

The open questions are reliability and accountability. A 93% success rate is high for a benchmark but low for a safety reading in a hazardous plant, and DeepMind's own model card warns against safety critical deployment. DeepMind has not published pricing for the preview, or said when it will be generally available.