Robotics‑ER 1.6 update

Robots need two kinds of intelligence: one to see and reason about a scene, and another to move through it. On April 14, Google DeepMind updated the first part with Gemini Robotics‑ER 1.6, a model built to help robots interpret physical spaces more precisely. (deepmind.google) Google said version 1.6 improves spatial reasoning, multi-view understanding, task planning, and “success detection,” which is the model’s check on whether a job was actually completed. The company said the model outperformed both Gemini Robotics‑ER 1.5 and Gemini 3.0 Flash on internal evaluations for pointing, counting, and physical reasoning. (deepmind.google) The new release also adds instrument reading, so a robot can interpret pressure gauges and sight glasses from camera images. Google said it developed that use case with Boston Dynamics and made the model available on April 14 through the Gemini application programming interface and Google AI Studio. (deepmind.google) Google introduced Gemini Robotics and Gemini Robotics‑ER in March 2025 as separate layers for physical machines. Gemini Robotics is the vision-language-action model that turns images and instructions into motor commands, while Gemini Robotics‑ER is the reasoning layer for perception, planning, and tool use. (deepmind.google) That split matters in factories and warehouses, where a robot may already know how to walk a route but still miss what it is looking at. Boston Dynamics said industrial inspections often require more than object detection, including checks for conveyor damage, sight-glass levels, pallet counts, spills, and gauge readings. (bostondynamics.com) Boston Dynamics said on April 15 that it is integrating Gemini and Gemini Robotics‑ER 1.6 into Orbit AIVI-Learning, the company’s cloud inspection system for Spot. The company said Spot and Orbit can now “continuously learn” about a facility, and that AIVI-Learning now supports gauges alongside 5S compliance audits, sight-glass measurement from 0 to 100 percent, pallet counting, and puddle detection. (bostondynamics.com) Boston Dynamics’ support documents describe AIVI-Learning as a cloud-hosted, multi-model system that combines Boston Dynamics “expert” models with Google Gemini. The company said customers must opt in to data sharing with Boston Dynamics to use AIVI-Learning, while Google “does not have access” to that shared inspection data for training its own models. (support.bostondynamics.com) Google is framing 1.6 as a reasoning upgrade rather than a full robot controller. Its product page says Gemini Robotics models can perceive, reason, use tools, and interact across different robot forms, from fixed arms to humanoids, while the ER model is the high-level layer that specializes in understanding physical spaces and making decisions inside them. (deepmind.google) Google also said Gemini Robotics‑ER 1.6 is its safest robotics model so far on adversarial spatial reasoning tests, a benchmark for whether the system follows safety rules when prompts or scenes get tricky. The immediate test for that claim will be whether robots using the model can keep turning camera feeds into reliable readings in working industrial sites, not just demos. (blog.google)