Gemini Robotics ER‑1.6 update
DeepMind published Gemini Robotics ER‑1.6, a model update that improves spatial reasoning and multi‑view understanding to help robots read instruments and perform real‑world tasks. Google’s own blog frames the update as an advance in embodied reasoning for autonomous systems operating in physical environments. (deepmind.google) (blog.google)
Google DeepMind published Gemini Robotics-ER 1.6 on April 14, adding stronger spatial reasoning, multi-view scene understanding, and a new ability for robots to read gauges and sight glasses. (deepmind.google) In plain terms, embodied reasoning is the part of robotics that lets a machine connect what its cameras see to what its body should do next. DeepMind said ER-1.6 is a high-level reasoning model for visual understanding, task planning, and checking whether a job is actually finished. (deepmind.google) The update is aimed at messy physical settings, not just lab demos. DeepMind said the model improved on Gemini Robotics-ER 1.5 and Gemini 3.0 Flash on internal tests for pointing, counting, and success detection, and it can call tools including Google Search, vision-language-action models, and user-defined functions. (deepmind.google) One of the new pieces is instrument reading, which means a robot can interpret analog gauges and liquid-level windows instead of just spotting objects. DeepMind said it developed that use case with Boston Dynamics. (deepmind.google) Boston Dynamics announced the same day that it is integrating Gemini and Gemini Robotics-ER 1.6 into Orbit AIVI-Learning for Spot, its quadruped robot used in industrial inspection. The company said the system will handle tasks including gauge checks, sight-glass measurements from 0 to 100 percent, pallet counting, and spill detection. (therobotreport.com) Google also said ER-1.6 is its safest robotics model so far. In its public summary, the company said the model showed better compliance with safety policies on adversarial spatial reasoning tasks, though the benchmark details released publicly are limited to Google’s own materials. (blog.google) This update builds on a push Google DeepMind started on March 12, 2025, when it introduced Gemini Robotics and Gemini Robotics-ER as Gemini 2.0-based models for physical robots. In that launch, DeepMind positioned ER as the reasoning layer and paired the broader robotics effort with partners including Apptronik and trusted testers such as Boston Dynamics. (deepmind.google) The thread running through both launches is that robot makers want a system that can handle new objects, new camera angles, and incomplete information without being retrained for every edge case. ER-1.6 does not turn a robot into a finished product on its own, but it gives developers a new planning and perception layer through the Gemini Application Programming Interface and Google Artificial Intelligence Studio starting April 14. (deepmind.google)
