Gemini expands into robotics

Google DeepMind has moved Gemini deeper into robotics, releasing Gemini Robotics-ER 1.6 on April 14 and making it available through the Gemini application programming interface and Google AI Studio. (deepmind.google) A robotics model is the planning layer that turns camera views and language into steps a machine can carry out in the physical world. Google said ER 1.6 is a “reasoning-first” model for visual and spatial tasks such as counting objects, planning actions, and checking whether a job is finished. (deepmind.google, ai.google.dev) Google said the new version improves spatial reasoning and multi-view understanding, which means a robot can use more than one image or angle to judge what is happening around it. The company also said the model can call outside tools, including search, vision-language-action systems, or other user-defined functions while planning a task. (deepmind.google) The new feature Google highlighted most is instrument reading: robots can now interpret pressure gauges and sight glasses in industrial settings. Google said that use case came from work with Boston Dynamics, which builds the four-legged Spot inspection robot. (deepmind.google, bostondynamics.com) Boston Dynamics said on April 15 that it is integrating Gemini and Gemini Robotics ER 1.6 into Orbit AIVI-Learning, its cloud system for artificial-intelligence visual inspections. The company said Spot and Orbit will use the models for higher-order reasoning and more complex visual analysis inside facilities. (bostondynamics.com) That Spot announcement builds on a broader Google DeepMind and Boston Dynamics tie-up unveiled at the Consumer Electronics Show on January 5, 2026. At that event, the companies said they would use Gemini Robotics models in joint research on Boston Dynamics’ new Atlas humanoid robot, with work aimed first at industrial tasks in automotive manufacturing. (bostondynamics.com) Boston Dynamics already sells Spot as a mobile robot for inspections in factories, plants, and other industrial sites, while Orbit is the software layer that manages fleets and analyzes inspection data. Its support documents say AIVI-Learning is a cloud-hosted, multi-model system that includes Boston Dynamics models plus Google Gemini and improves over time as new data is applied. (bostondynamics.com, support.bostondynamics.com, bostondynamics.com) Google’s earlier public robotics documentation for ER 1.5 described the model as a vision-language system that can locate objects, understand relationships, and break natural-language instructions into subtasks. ER 1.6 extends that line with stronger physical-world reasoning and a new industrial reading task that links the model more directly to deployed robots rather than chatbots or coding assistants. (ai.google.dev, deepmind.google) Google also said ER 1.6 is its “safest robotics model to date,” citing stronger compliance with safety policies on adversarial spatial reasoning tests. In the company’s framing, the model sits at the top of a robot stack: it reasons about a scene, then hands off actions to controllers and other tools that actually move the machine. (blog.google, deepmind.google) The immediate next step is practical rather than theatrical: more robots reading meters, spotting anomalies, and planning inspection work inside industrial sites. Google has opened the model to developers, and Boston Dynamics has started wiring it into the software that already runs Spot fleets in the field. (deepmind.google, bostondynamics.com)