Reading the pressure on a factory gauge sounds simple, but until now robots couldn't do it.
It's not that robots can't see the dial — they can take a picture and recognize a circular instrument. But making a judgment like "the needle points to 237, the unit is PSI, and the reading is abnormal" requires combining visual understanding, numerical reasoning, contextual knowledge, and knowing what "237 PSI" means for that specific piece of equipment.
On April 15, Google DeepMind released Gemini Robotics-ER 1.6, providing a concrete number for this capability: gauge reading accuracy of 86%, rising to 93% with Agentic Vision enabled. Compared to the previous version? 23%.
This is not a minor iteration. It's a leap from "basically unusable" to "ready for industrial deployment."
What the model actually does
Gemini Robotics-ER stands for Embodied Reasoning — enabling AI not just to see, but to plan actions and determine whether a task is complete based on what it sees.
Version 1.6 upgrades four main areas:
Instrument Reading
This is the most critical new capability. When Spot robots inspect factories, they face analog gauges like pressure meters, thermometers, and level indicators that require precise reading. Previous vision AI often failed between "the gauge is around the middle" and "it's exactly 237 PSI."
Version 1.6 combines visual reasoning with code execution: it captures an image, uses code to estimate the dial range and needle position, and then outputs a reading — with accuracy to less than one scale division. This was a need identified by DeepMind and Boston Dynamics during real-world use, and they developed it specifically.
Multi-View Understanding
Robots typically have multiple cameras. Earlier models often got confused when processing multiple video feeds — "which camera is this object in? Is it the same object in the other camera?" Version 1.6 can simultaneously process multiple visual inputs and integrate them into a coherent scene understanding.
Pointing & Spatial Reasoning
By first locating key points and then reasoning about spatial relationships, the model improves accuracy in tasks like counting objects and finding grasp points.
Success Detection
Determining whether a task has been completed in occluded or poorly lit environments has always been a challenge for robots. Version 1.6 shows significant improvement here, allowing robots to autonomously decide whether to retry — without needing to call a human remotely.
How Boston Dynamics is using it
Boston Dynamics' Spot robot has a mature industrial inspection product line: entering factories for patrols, recording audio and video, and logging equipment status. But previously Spot could see the gauge dial without accurately reading the numbers — either a worker had to follow along, or Spot would take photos and send them back for manual interpretation.
Now that step can be eliminated.
Boston Dynamics VP Marco da Silva said: "Instrument reading and more reliable task reasoning will allow Spot to fully autonomously see, understand, and respond to real-world challenges."
This is not marketing talk — it's a concrete change in use case: instead of a person following Spot to record each device's reading, Spot can now complete an entire inspection route autonomously, and the person only needs to review the final report.
The numbers tell the story
Direct performance comparison:
| Capability | Gemini Robotics-ER 1.5 | Gemini Robotics-ER 1.6 |
|---|---|---|
| Gauge reading accuracy | 23% | 86% (Agentic Vision: 93%) |
| Video safety risk perception | Baseline | +10% |
| Text safety risk perception | Baseline | +6% |
The jump from 23% to 86% in gauge reading essentially means this feature has moved from prototype to practical use.
What the API opening means
Gemini Robotics-ER 1.6 is available to developers through the Gemini API and Google AI Studio. This distribution method is significant.
Hardware manufacturers (not just Boston Dynamics) can now directly call this model to integrate advanced visual reasoning into their own robot platforms — no need to train their own models or maintain model weights, just an API call. This dramatically lowers the barrier and allows DeepMind's robot AI to penetrate various industrial scenarios more quickly.
This differs from Waymo's logic of accumulating road test data as a moat: DeepMind is taking a platform approach, letting third-party hardware partners proliferate the capability. Any robot that uses it gains a directly commercializable capability in industrial inspection.
Industrial robots may still be some distance from "not needing a single person to follow in the factory," but Spot being able to read gauges has already brought that point much closer.
Sources: Gemini Robotics ER 1.6: Enhanced Embodied Reasoning (Google DeepMind Blog); DeepMind launches Gemini Robotics-ER 1.6 to meet precise physical AI demands (SiliconANGLE); CocoLoop; Google DeepMind Releases Gemini Robotics-ER 1.6 (MarkTechPost)