← Back to feed

Google DeepMind released Gemini Robotics-ER 1.6 on April 14, an updated version of its Embodied Reasoning foundation model that acts as the high-level strategic planner in DeepMind’s two-tier robot brain architecture. The model is available today via the Gemini API and Google AI Studio.

The release focuses on three capability areas where ER 1.5 showed meaningful gaps: pointing-based spatial reasoning, multi-view success detection, and a new instrument-reading function. All three are requirements for robots operating in industrial environments where a camera feed alone is insufficient — the robot must interpret what it sees and act on it autonomously.

What Changed

Pointing and spatial reasoning. The updated model uses points as intermediate reasoning steps to identify objects, count items, map movement trajectories, and locate grasp positions within a scene. It can also correctly decline to point when the target object is not present — reducing false-positive hallucinations that caused downstream errors in ER 1.5.

Multi-view success detection. Determining whether a task has been completed is a core challenge for autonomous robots. ER 1.6 can synthesise information from multiple simultaneous camera streams — such as a ceiling-mounted wide-angle camera and a wrist-mounted feed — to determine task completion even in partially occluded or dynamically changing environments. Earlier versions handled this poorly when views conflicted.

Instrument reading. The headline new capability was developed through collaboration with Boston Dynamics. ER 1.5 achieved a 23% success rate reading physical gauges and sight glasses; Gemini 3.0 Flash reached 67%. ER 1.6 shows substantial improvement over both, according to DeepMind’s internal benchmarks, though the company has not published a precise figure. The use case is significant for manufacturing, utilities, and facility management — environments where Boston Dynamics’ Spot robot is already deployed and where reading analogue instruments without human oversight is a recurring requirement.

Architecture Context

In DeepMind’s robotics stack, the ER model functions as the brain’s reasoning layer. It interprets open-ended goals, calls tools (including Google Search and third-party APIs), and directs Vision-Language-Action models — the motor cortex — to execute physical movements. ER 1.6 extends that tool-calling capability to allow natively invoking external VLAs and user-defined functions, widening its integration surface.

DeepMind describes ER 1.6 as its safest robotics model to date, citing improved compliance with safety policies on adversarial spatial reasoning tasks — a relevant claim as industrial deployments move toward less-supervised operation.

The rapid iteration pace — ER 1.5 shipped in late 2025 — signals that DeepMind is treating robotics as a near-term production frontier rather than a research programme. Developer access via the standard Gemini API, rather than a separate research SDK, reinforces that positioning.