Google DeepMind's Gemini Robotics-ER 1.6 Brings AI Reasoning to Boston Dynamics' Spot

For years, the gap between robot hardware and robot intelligence has been the defining problem of industrial automation. The hardware — cameras, actuators, sensors — has gotten extraordinarily capable. The intelligence layer has lagged, requiring robots to operate within narrowly scripted environments with little tolerance for the messiness of real-world conditions.

On April 15, 2026, Google DeepMind released Gemini Robotics-ER 1.6, and that gap got meaningfully smaller.

The model, described by DeepMind as a “reasoning-first” upgrade to its Gemini Robotics line, brings a qualitatively different kind of intelligence to physical systems: the ability to look at an environment and actually understand it — not just detect objects, but reason about spatial relationships, assess task completion, and decide when to escalate to more specialized tools. On the same day, Boston Dynamics announced it is integrating Gemini Robotics-ER 1.6 directly into the Orbit software platform that powers its Spot robot’s industrial inspection workflows, marking one of the most concrete physical AI deployments by a major robotics company to date.

The model is available immediately to developers via the Gemini API and Google AI Studio, with no waitlist.

What “Embodied Reasoning” Actually Means

The “ER” in Gemini Robotics-ER stands for Embodied Reasoning — a deliberately specific framing that distinguishes this model from conventional vision-language models.

Standard vision-language models are good at describing what they see. Embodied reasoning models are designed to understand what to do with what they see — a distinction that sounds subtle but matters enormously in physical environments. A robot equipped with a description-only model can tell you there is a gauge on the wall. An embodied reasoning model can read the gauge, determine whether the reading is within normal operating parameters, and decide whether the situation requires immediate action.

Gemini Robotics-ER 1.6 specifically advances three capabilities that have been bottlenecks for industrial robotics:

Spatial reasoning: The model can interpret multi-camera views and construct accurate mental models of three-dimensional environments. This enables a robot to determine object positions, understand depth and occlusion, and navigate complex scenes without pre-mapped environments.

Task completion assessment: Rather than relying on scripted success criteria, the model can evaluate whether a task has been accomplished by interpreting visual evidence — a critical capability for autonomous quality inspection workflows where success is often ambiguous from a single frame.

Tool escalation: When the model encounters a situation beyond its direct capabilities, it can identify and invoke specialized vision-language-action (VLA) models to handle the specific subtask. This meta-capability turns Gemini Robotics-ER 1.6 into an orchestration layer for a broader ecosystem of specialized robotics models.

Boston Dynamics’ Bet on AI-Powered Inspection

The partnership with Boston Dynamics is the centerpiece of the Gemini Robotics-ER 1.6 launch, and it is immediate and production-focused rather than experimental.

Boston Dynamics is integrating the model into two components of its Orbit software platform: AI Visual Inspection (AIVI), which handles real-time anomaly detection during Spot’s patrol routes, and AIVI-Learning, which enables the system to improve its detection capabilities over time from operational data.

With Gemini Robotics-ER 1.6 onboard, Spot gains a set of capabilities that mark a qualitative leap for industrial inspection robots. The system can now autonomously detect dangerous debris or liquid spills on floors and surfaces — a safety-critical function in manufacturing, oil and gas, and utility environments. It can read complex gauges and sight glasses, interpreting analog instruments that previously required human visual assessment. And it can call on specialized VLA models when it encounters something outside its direct knowledge — a manufacturing anomaly it hasn’t seen before, an unfamiliar equipment configuration, or a situation requiring fine manipulation reasoning.

Industrial inspection has been one of the most compelling early applications for autonomous mobile robots, but the limiting factor has consistently been intelligence rather than mobility. Spot can already navigate complex environments, manage stairs, and operate in hazardous conditions where human inspectors face risk. What it has lacked is the ability to make nuanced judgments about what it observes. Gemini Robotics-ER 1.6 directly addresses that gap.

The Physical AI Race Heats Up

The release lands in an increasingly competitive landscape for physical AI. NVIDIA has been investing heavily in its own physical AI platform, with Cosmos for world model simulation and Isaac GR00T for robot learning. Microsoft has robotics ambitions through its Azure infrastructure. Amazon has made major moves in warehouse automation. And a wave of humanoid robot startups — Figure, Physical Intelligence, 1X, Neura Robotics — are building vertically integrated systems that combine custom hardware and proprietary AI.

What Google DeepMind is doing with Gemini Robotics-ER is a different play: a horizontal model that runs on existing hardware — quadrupeds, arms, mobile bases — across manufacturers and applications. Rather than owning the robot, DeepMind aims to own the intelligence layer.

This is consistent with Google’s broader strategy. Gemini already runs across phones, laptops, cloud services, and now robots. Making Gemini Robotics-ER available via API means any developer building robotics applications can call on DeepMind’s latest research without needing to build and train embodied reasoning capabilities from scratch. It also means Google collects deployment data at scale — a critical ingredient for improving subsequent model generations.

The Boston Dynamics partnership is particularly notable because Boston Dynamics occupies a unique position in the robotics market: it makes hardware that people genuinely trust with safety-critical tasks. Its Spot robots are deployed in nuclear power plants, offshore oil platforms, construction sites, and disaster response environments. Having Gemini Robotics-ER 1.6 running in those deployments gives DeepMind real-world feedback from conditions that no simulation can replicate.

Developer Access and What Comes Next

Unlike many frontier model releases that launch through exclusive programs before broader availability, Gemini Robotics-ER 1.6 is available to developers immediately via the Gemini API and Google AI Studio. DeepMind’s decision to release openly reflects both competitive pressure — NVIDIA’s Cosmos and GR00T are gaining traction, and the physical AI talent pool is being contested aggressively — and a genuine belief that ecosystem breadth matters more than controlled access in this phase of the market.

DeepMind has indicated that subsequent versions will continue the trajectory toward more general embodied intelligence: models that can learn from a single demonstration, generalize across robot morphologies, and plan multi-step manipulation sequences without explicit programming. The company has also been working on tighter integration between Gemini’s reasoning capabilities and robot hardware through the Gemini Robotics-Actions line, which handles low-level motor control and is expected to see its own update later in 2026.

For robotics developers and integrators, Gemini Robotics-ER 1.6 raises the floor of what’s achievable with off-the-shelf AI. Robots don’t need to be programmed for every scenario if the model can reason through novel situations. That shift — from scripted automation to adaptive intelligence — is the transition the industry has been building toward for years.

It’s no longer theoretical.