Google DeepMind brings thinking AI agents to robots

Sorting household trash sounds simple to a human, but it is a logistical nightmare for a robot. The machine must check local recycling rules online, visually inspect the items on the table, figure out which bin each item belongs in, and coordinate precise motor controls to pick them up without knocking anything over.

Google DeepMind unveiled two models built to tackle this challenge: Gemini Robotics 1.5 and Gemini Robotics-ER 1.5. Together, they divide physical robotics tasks into high-level strategic reasoning and low-level physical execution.

Developers can start testing Gemini Robotics-ER 1.5 right now in preview through Google AI Studio and the Gemini API. The motor control model, Gemini Robotics 1.5, is currently rolling out to select partners.

Splitting the brain from the hands

Instead of trying to run physical manipulation and open-ended planning through a single pipeline, Google split the architecture in two.

Gemini Robotics-ER 1.5 acts as the high-level brain. It excels at spatial understanding, evaluating video streams, and estimating task progress. When asked to complete a contextual task, it can natively query Google Search to check real-world information or call custom user-defined APIs before issuing instructions.

In practical developer demos, the model outputs normalized 2D coordinates for target objects in an image. When paired with depth sensors, this gives robots precise 3D spatial grounding. It can also track state changes over time, like identifying whether a laptop is open or closed, and break down video recordings of physical tasks into timestamped event sequences.

Once ER 1.5 finishes mapping out the sub-tasks, it passes clear instructions down to Gemini Robotics 1.5.

Thinking before grabbing

Traditional vision-language-action (VLA) models convert text and visual input directly into joint commands. If anything unexpected happens, the execution collapses.

Gemini Robotics 1.5 adds a reasoning buffer. It thinks in natural language before executing, generating an internal chain of reasoning to break long actions into smaller segments. When asked to sort laundry, for example, it reasons about which items belong in which bin and decides whether to drag a sweater closer before attempting a lift.

DeepMind also reported cross-embodiment learning. Skills demonstrated solely on an ALOHA 2 dual-arm setup transferred successfully to Apptronik’s humanoid robot Apollo and a Franka arm without retraining for the new physical platforms.

On the safety side, Google updated its ASIMOV benchmark, integrating semantic risk assessments into the model’s planning cycle to trigger local collision prevention sub-systems when hazards arise.

Why it matters

For robotics developers, open access to an embodied reasoning model via API changes the game. Building spatial grounding from scratch is prohibitively expensive; now, small teams can plug ER 1.5 into custom hardware using Google AI Studio and run multi-step planning off the shelf.

Across the industry, the cross-embodiment transfer is the standout technical achievement. If training data gathered on one dual-arm setup reliably transfers to full humanoid frames, physical robotics will overcome one of its worst data bottleneck problems.

We think Google’s two-tier approach is pragmatic. Decoupling heavy semantic planning from immediate motor execution gives developers transparency and flexibility. The real hurdle ahead, however, is real-time latency: coordinating API-based reasoning calls with sub-second physical reaction loops in messy real-world environments remains a tricky balancing act.